Anthropic Startup Program: Claude Credits Guide 2026

What Anthropic's startup program actually offers, who qualifies for Claude credits, and every other place a startup can get its Claude bill paid in 2026.

Anthropic's startup program page does not state a credit amount anywhere. The Claude for Startups page promises venture-backed founders "free credits and priority rate limits", sets three conditions for them (equity funding from an institutional investor, a company founded within the last four years, and no earlier Anthropic startup credits), and stops there. The figures that circulate online, tiers of $5,000, $25,000 or $100,000, come from third-party sites, not from any Anthropic page we could find on October 5, 2026. Until today our own catalog repeated one of them as "up to $100K"; we corrected it while researching this guide, and the correction is the reason this guide exists.

The second surprise is where Claude credits can be spent. Anthropic's own FAQ says its startup credits "can't be used on AWS Bedrock, Google Cloud Vertex AI, or other third-party platforms" - Anthropic. Yet the largest pool of credit that can pay for Claude is not Anthropic's at all: AWS lets its Activate credits pay for third-party models on Amazon Bedrock and names Anthropic among them - AWS Startups FAQ. Google's and Microsoft's startup credits, by contrast, exclude Claude in writing. Which cloud you choose decides whether your cloud credits buy Claude tokens, and most founders find that out after the bill arrives.

This guide lays out every route we verified on its owner's own page on October 5, 2026: Anthropic's startup program and its partner-VC track, the Menlo Anthology Fund, accelerators that bundle Claude credits, hackathons and events, Anthropic's research programs, the three big clouds, model gateways, platform credits and banking perks. It explains how the Claude Console's usage tiers and monthly spend caps interact with a grant (a new organization on the Start tier can spend at most $500 a month), prices the credits in real requests, and ends with a decision framework, the order to apply in and the mistakes that waste credits. Every number links to the page it came from.

Contents

1. The short answer: every route at a glance
2. How Claude API credits work: the Console, tiers and spend caps
3. Route 1: Claude for Startups, Anthropic's own program
4. Route 2: partner VCs and the Anthology Fund
5. Route 3: accelerators that bundle Claude credits
6. Route 4: hackathons, events and community programs
7. Route 5: research, education, nonprofit and open-source programs
8. Route 6: AWS Activate and Claude on Amazon Bedrock
9. Route 7: Google Cloud, where startup credits stop at Gemini
10. Route 8: Microsoft Foundry, where Claude bills the card
11. Route 9: gateways, platforms and banking perks
12. What Claude credits buy: the burn math
13. Making credits last: caching, batching and model choice
14. Which route fits your startup
15. Stacking routes, and the order to apply
16. Eight mistakes that waste Claude credits
17. Frequently asked questions
18. How we verified this guide

1. The short answer: every route at a glance

There are two kinds of credit that pay for Claude, and they are not interchangeable. The first kind is Anthropic's own: a balance in your Claude Console organization, spent through the first-party Claude API. The second kind is another company's credit that happens to pay for Claude when you call it through that company's platform: AWS Activate credits on Amazon Bedrock, OpenRouter's inference credits, Vercel's credit on its AI Gateway. Both buy the same models, and in most cases at the same list price. Only the first is what people mean by "Claude credits", and it has the narrowest door.

The table below lists every route we verified, what its owner publishes about the amount, who qualifies and where the credit is spent. Where an owner publishes no amount, the table says so instead of repeating a figure from someone else's blog. Programs with a page in our catalog link to it; the rest link to the owner.

Read the table from the right-hand column first. If your product has to call the first-party Claude API, for example because you depend on a feature that ships there first or on the Console's tooling, only the rows that end in "Claude API" help you. If you can call the same models through Amazon Bedrock, the AWS rows open budgets that are far larger and far easier to reach, because AWS Activate does not require venture funding at its first tier.

The chart makes one structural point plain. The biggest single pool of credit that can pay for Claude is AWS's, and it is general-purpose cloud credit: every dollar spent on Claude through Bedrock is a dollar not spent on compute, storage or databases. Vercel's bar is half of its $30,000 award, because Vercel caps the share that can go to AI Gateway at 50%. Anthropic's own program cannot appear on the chart at all, because it publishes no number.

Why the map looks like this comes down to who bills the tokens. A startup credit is a customer-acquisition subsidy, and a company only subsidizes spend it collects. On Bedrock, AWS is the invoicing party for third-party models - AWS Service Terms, so AWS can choose to let its own credits absorb the cost of Claude, and it does, because a startup that runs its models on Bedrock tends to run everything else on AWS too. On Microsoft Foundry, Anthropic is the seller of record for Claude, and Microsoft's rule is blunt: "If Microsoft bills the AI model usage to your Azure subscription, Microsoft for Startups sponsorship credits apply" - Microsoft Learn. Claude is not billed that way, so the credits do not apply. Google draws the same line around "third-party models". Coverage follows the invoice, not the model, and once you see that, every rule in this guide becomes predictable.

Anthropic's own credits follow the same logic from the other side. Anthropic does not need to buy demand: its annualized revenue run rate reached $65 billion at the end of July, according to CNBC. Its startup credits are therefore rationed by signal rather than by budget. The application asks for your total current LLM spend across all providers and your lead investors, because both predict how large a customer you will become. That is why the door is narrow, and why the strongest application is a company that already runs Claude in production, which section 3 turns into a practical checklist.

2. How Claude API credits work: the Console, tiers and spend caps

Every Anthropic credit lands in one place: a Claude Console organization. The Console is Anthropic's developer platform, separate from the Claude chat app. A paid Claude subscription "doesn't include access to the Claude API or Console" - Claude Help Center, and a Claude account and a Console account on the same email address "operate independently" - Claude Help Center. Founders who pay for Claude Max and expect their startup credits to show up in the chat app are looking in the wrong place: credits fund API calls, the Console playground and, when you sign in with a Console account, Claude Code.

The Console runs on prepaid usage credits. Anthropic's billing article says credits "cover API access, playground usage, and Claude Code", that purchased credits "expire one year from the purchase date", that the expiration "can't be extended", and that all purchases are non-refundable - Claude Help Center. Only successful calls are billed, the API stops when the balance hits zero, and an auto-reload option tops the balance up when it falls below a threshold you set. Free credits follow the Supplemental Credit Terms: promotional credits "expire at the time indicated when issued (or otherwise one calendar year from when issued if no time is specified)", cannot be transferred or sold, and vanish if you close the account - Anthropic.

A new account gets a little free credit, and nobody official says how much. The pricing documentation says only that "New users receive a small amount of free credits to test the API" - Claude Docs. The $5 figure repeated across the web is not on any Anthropic page we read. Treat sign-up credit as a smoke test, not a budget.

Usage tiers and the monthly spend cap

The tier system changed in 2026, and most guides still describe the old one. There are now three usage tiers, Start, Build and Scale, plus a Custom tier for accounts managed by an account team - Claude Help Center. The old numbered tiers with deposit thresholds are gone: "There's no deposit or purchase that moves you up" - Claude Help Center. Organizations are placed automatically "based on usage history and account standing", brand-new ones may start in a lower Evaluation tier whose limits are not published, and you can request higher limits in the Console once you use at least half of your current ones.

Each tier also carries a monthly spend cap, and this is the detail that matters most for anyone holding a grant. Anthropic's rate-limit documentation sets the caps at $500 for Start, $1,000 for Build and $200,000 for Scale, with no cap on Custom - Claude Docs. When an organization reaches its cap, "API usage pauses until 00:00 UTC on the first day of the next month, unless you request a higher limit sooner." Credits do not lift the cap. A startup that receives a five-figure grant while sitting on the Start tier can still only spend $500 of it a month.

The table also explains a line on the startup program's FAQ that is easy to skim past: "Acceptance upgrades you to our highest rate limits, so you can ship to production without throttling while you're on credits" - Anthropic. The page does not name the tier, but the upgrade is worth as much as the credit itself, because without it a large grant would take years to spend at the Start tier's pace.

The per-model limits scale with the tier. These are the published standard limits for the current models, in requests per minute (RPM), input tokens per minute (ITPM) and output tokens per minute (OTPM) - Claude Docs:

Caching raises your effective throughput, not only your savings. "For most Claude models, only uncached input tokens count toward your ITPM rate limits", and Anthropic's own example notes that with a 2,000,000 ITPM limit and an 80% cache hit rate, you could process 10,000,000 input tokens a minute. Section 13 returns to caching; for now, note that the same habit that stretches a grant also stretches a rate limit.

Two more rules round out the picture. Claude Code, when used with a Console account, "consumes API tokens at standard API pricing" - Claude, so a startup's API credits can fund its engineers' coding agent as well as its product. And Anthropic is rolling out identity and business verification "for a few use cases", which can prompt a check when you access certain capabilities - Claude Help Center. Neither changes what a credit is worth, but both change how quickly you can use one, so set up the Console organization, its billing and its verification before the credits arrive rather than after.

How to apply this section: open the Console organization you will build in, note its API Organization UUID (the startup application asks for it), check the tier and spend cap on the Billing and Limits pages, and request a limit increase as soon as real usage justifies it. A grant that lands in a well-configured organization starts working the same day; one that lands in a fresh Evaluation-tier account spends its first weeks throttled.

3. Route 1: Claude for Startups, Anthropic's own program

The program has two layers, and only one of them includes credits. The first layer is membership. "Any early-stage founder or startup building with Claude can join the Startups program", and Anthropic answers "Do I need VC funding to apply?" with "No. The program is open to any early-stage founder or startup with or without VC backing" - Anthropic. Members get community events, resources and early news. The second layer is the one people search for: "Venture backed? Apply for free credits and priority rate limits. Token max to product market fit on us."

The program page frames the benefits as three columns: credits, community, and staying close to Anthropic. What it leaves out is as informative as what it says. There is no dollar figure, no tier table, no expiry date and no list of partner investors anywhere on the page.

The credit layer has three conditions, stated in the FAQ: "your company must have received equity funding from institutional investor, been founded within the last four years, and not previously received Anthropic startup credits." Read literally, that excludes three groups of founders that third-party guides often say can apply: bootstrapped companies (no institutional equity), companies funded only by angels or friends (no institutional investor), and any company that has already received Anthropic startup credits once. The last condition makes this a one-time grant per company, which matters for timing: apply when the credits will carry you furthest, not on the day you incorporate.

What the application asks

The application is short. Anthropic says it "takes about two minutes" and needs "a Claude Console account (the API console, not the Claude app) and a company email address". The form itself, which we read without submitting, asks for more than the FAQ suggests - Claude for Startups application. The required fields include your API Organization UUID (so the credits attach to one specific Console organization), "What is your total current LLM spend across all providers?", "Where do you want support from Anthropic?", and a 50 to 500 character description of your startup and how you plan to use Claude.

Two fields are optional but carry weight. The form asks for your latest funding round, with options from Bootstrapped and Angel through Pre-Seed, Seed, Series A and Series B+, and "If you've raised, who are your lead investors?" The program page tells VC-backed founders to "Mention your investor when you apply", and adds that founders backed by "an Anthropic partner VC or accelerator" may qualify for additional benefits. Section 4 covers the partner track.

What the official terms add

Anthropic published Startup Program Official Terms, last updated May 22, 2025 - Anthropic. The current form does not link to them, so treat them as the last published rules rather than a guarantee of today's, but their wording is consistent with the program page. Selected companies "may receive Anthropic API credits, priority rate limits, and/or access to certain resources", attributed to the account of the company named in the application. Applicants are selected "on a rolling basis", and Anthropic weighs "business traction, investment, and funding, as well as Claude integration and usage", with eligibility "determined in Anthropic's sole discretion".

The terms also carry obligations that surprise founders. A selected company may have to return a signed affidavit and release within fourteen days or forfeit the benefits. Anthropic may use your name and logo to identify you as a selected company and customer. Credits "may have different tax implications than cash awards". And companies established or operating in Belarus, China, Cuba, Iran, Myanmar, North Korea, Russia, Sudan, Syria, Crimea or the so-called Donetsk and Luhansk People's Republics are not eligible.

What you get, and what happens after

Accepted startups get credits in an amount Anthropic does not publish, the highest rate limits while on credits, and the community layer: "hackathons, Founder Days, and meetups in six cities", livestreams and AMAs, and early notice of launches and model releases. When the credits end, "Standard API pricing kicks in automatically with no migration", and the startups team offers help optimizing costs. The program page states no expiry; under the general credit terms quoted in section 2, promotional credits run until the date stated when they are issued, or one year if no date is given.

Anthropic's short film from its Frontier Day event shows who the program is pitched at: founders speaking about building startups with Claude. It carries no figures, but it is the program's official video, and its description links straight to the program page.

How to apply well

Anthropic tells you what it weighs: traction, funding, and Claude integration and usage. The credit is a bet on your future spend, so make that bet easy to price. Everything on the form either proves you already use Claude, proves an institutional investor has checked your company, or tells Anthropic what its help would unlock. A reviewer comparing hundreds of two-minute applications has little else to go on, so the few fields you control carry the whole decision.

That reframes preparation. An application is not a pitch deck: there is no room for vision statements, and a 500-character description rewards one concrete sentence about the product and one about how Claude is used in it. The rest of the evidence is structural, and you can build it before you apply. Have five things in place:

• A working Console organization with real Claude usage, whose UUID you submit
• A company-domain email and a live website that explains the product
• Your institutional investor's name, ideally confirmed as an Anthropic partner
• An honest LLM spend figure, across every provider you use today
• A specific support ask, such as rate limits for a launch or architecture help

The order of that list is deliberate. The organization and its usage are the strongest evidence you can give, because Anthropic can see them: a startup that already sends production traffic through the Console it names is a lower-risk bet than one that promises to start. The investor's name is the second-strongest signal, because institutional funding is a hard condition and a partner relationship may unlock more. The spend figure and the support ask tell the reviewer what you will do with the grant. Do not inflate the spend figure: it is checkable once you are a customer, and the program's value comes from a long relationship, not one grant.

If you are bootstrapped or angel-funded, join anyway. Membership is open to you, the community events are where hackathon credits and introductions come from, and the rate-limit path still applies as your usage grows. Then take the AWS route in section 8, which is the one large credit program that pays for Claude without asking for a venture investor. Our Claude for Startups program page tracks the program's live terms, and our best AI credits for startups list ranks it against every other model-credit program.

4. Route 2: partner VCs and the Anthology Fund

The second way into Anthropic's own credits runs through your investors. The startup page carries a card for equity investors: "Investing in startups? You may be able to give founders access to free credits. VC funds can apply to join our program." The partner page spells out who Anthropic wants: "Venture capital firms that invest in AI-native startups across stages, with portfolio companies already building on or eager to adopt Claude" - Anthropic VC partner program. Anthropic says it weighs fund performance and market traction, the maturity of the fund's AI investment strategy, and existing Claude adoption among its portfolio companies. The application's fund-type list is broad: accelerators, seed funds, venture funds, private equity, university programs, venture studios, angel investors, coworking spaces, corporate partners and hackathon organizers can all apply.

What a partner receives is set out in the VC Partner Program Official Terms, last updated May 23, 2025: "Selected Awardees may receive Anthropic API credits and/or access to certain events or early updates for product releases for their portfolio companies" - Anthropic. Anthropic publishes no list of partner funds. The logo wall on the startup page shows customers such as Cursor, Lovable, Replit and Harvey, not investors, so the only reliable way to know whether your investor is a partner is to ask them.

Ask that question explicitly, and early. The startup application lets you name your lead investors, and the FAQ says founders "backed by an Anthropic partner VC or accelerator" should mention it "and you may qualify for additional benefits". If your investor is a partner, they can tell you whether the benefit comes through the standard application or through a separate route for their portfolio. If they are not, the standard application still applies to you as long as the investor is institutional, which is the condition that matters.

The Menlo Anthology Fund

The only startup credit amount Anthropic states on its own site is attached to an equity fund. In July 2024 Anthropic and Menlo Ventures launched the Anthology Fund, "a $100 million initiative financed by Menlo", and Anthropic's announcement says startups it backs gain "$25,000 in free credits towards our most advanced models" - Anthropic. Menlo's fund page still advertises the same credit today, and adds that the fund invests "from seed to expansion stages, with investments starting at $100,000 (and up!)" - Menlo Ventures.

Read that for what it is: an investment that comes with credits, not a credit program you apply to. For a company Menlo wants to fund, $25,000 in Claude credits is a pleasant extra on top of a check of at least four times that size. For a company that only wants credits, it is the wrong door, because the price of entry is equity.

The structural reason Anthropic routes credits through investors is selection. A partner fund has already done the diligence Anthropic would otherwise have to do, and a funding round is the best available predictor of future API spend. Routing credits through funds lets Anthropic subsidize exactly the companies most likely to become large customers, at almost no review cost. For a venture-backed founder that makes the partner track the strongest lever available, and the founder's job is to make sure the investor actually pulls it.

5. Route 3: accelerators that bundle Claude credits

Most of the named Claude credit amounts outside Anthropic's own pages sit inside accelerator perk bundles. Accelerators differ from the startup program in one fundamental way: the credits are a side benefit of an investment, and the investment is the price. Every amount below is quoted from the program's own page as we read it on October 5, 2026. Where a program names Anthropic inside a larger bundle without saying how much of it is Claude credit, the table quotes the bundle and says so.

Two of the programs were equity-free, and both were time-boxed cohorts that have closed. The rest invest in your company on their own terms, and the Claude credit is a small fraction of what they offer. That changes the question from "which accelerator has the most Claude credit" to "which accelerator would I join anyway, and does it happen to include Claude credit".

Only two rows put a number on the Anthropic part alone, Sequoia Arc and AI Grant, both at $30,000. AI Grant's page also says that applications for its Batch 4 are closed, so treat its perk list as the most recent one it published rather than a promise for the next batch. Entrepreneur First, South Park Commons, Conviction and speedrun pool Anthropic with other providers, so their headline figures describe the whole bundle and your Claude share of it is not stated. Before joining any of them for the credit, ask the program what the Anthropic portion is and whether it comes through Anthropic's partner track, which would make it a one-time grant under the same "no earlier Anthropic startup credits" rule.

The two equity-free cohorts show how Anthropic places credits when it wants a group of companies building on Claude. The Agentic AI Accelerator ran from May to September 2026 for 40 startups building agentic products from India, with "$50K AWS credits, valid for 12 months", "$20K Anthropic credits, valid for 6 months", access to Anthropic's engineers and an in-person kickoff in Bengaluru on June 26, 2026; its site runs under its own name rather than an Anthropic domain, and it lists applications as closed. The Vercel AI Accelerator opened its 2026 cohort in February with applications due February 16 and listed $600,000 in Anthropic API credits among more than $6 million in partner credits. Both paired Anthropic with a cloud or platform partner, both were limited to 40 teams, and both had an end date, so the right move is to watch for the next edition rather than wait for a permanent program.

How to apply this: never choose an accelerator for its credits. South Park Commons' $400,000 for 7% implies a post-money valuation of about $5.7 million, and the Claude credits inside its bundle are a rounding error against that. Choose on the people, the network and the terms, then collect the credits as a bonus, and remember that a credit received through an accelerator may count as your company's one Anthropic grant. Our accelerators directory tracks cohort dates and terms for these and many other programs, and our analysis of what 1,076 startup perk programs reveal shows how often credits come bundled with an investment rather than offered on their own.

6. Route 4: hackathons, events and community programs

Hackathons are the most open door to Anthropic's own credits. No investor, no funding round and no company age limit: the selection rewards a convincing project idea. Anthropic's virtual Claude Code hackathons, run with Cerebral Valley, follow a consistent pattern: "We'll select 500 participants and give each one $500 in Claude API credits to build for a week", with participants competing for "$100k in Claude API credits" - Cerebral Valley. Anthropic's own write-up of one of these events describes the same structure, with judges from the Claude team picking six winners from "a total prize pool of $100,000 in Claude API credits" - Claude blog.

In person, the format scales up. At the Claude Build Day in San Francisco on June 13, 2026, "More than 1,500 people had applied; 310 took part, many traveling from around the world, each with $500 in credits and one day to turn an idea into a working demo" - Claude blog. The event's listing adds "$150K in Claude credits across 3 finalists", judged with investors from Neo, a16z, Pear VC and others in the room - Cerebral Valley. A life-sciences edition run with the Gladstone Institutes gives each of 500 participants "a month of Claude Max 20x and $200 in API credits" - Cerebral Valley.

University Claude Builder Clubs run smaller events with real credit prizes. The Columbia and NYU hackathon on April 12, 2026 offered "$3,000 in Anthropic credits + cash prizes" and $25 in API credits to eligible students - Devpost. At USC, first place won $2,500 in Anthropic API credits and second place $1,500, and participants were told to have "their credits and API keys set up at console.anthropic.com" - Devpost. For a founding team still at university, these are a realistic first few thousand dollars of Claude credit.

Founder events are the other half of this route, and they are about relationships rather than credit. Anthropic's events page can be filtered to startup events and lists Claude Founder House gatherings in San Francisco on October 6, 2026, Stockholm on October 14 and London on September 23 - Anthropic events. None of those listings states credits for attendees. The Claude Community Ambassadors program does: city organizers "Get monthly API credits from Anthropic to power your demos", with no amount stated, and applications are open - Claude.

What $500 is worth, and how to use this route. At list price, $500 buys 250 million input tokens or 50 million output tokens on Claude Sonnet 5.5, which costs $2 and $10 per million - Claude Docs. It also happens to equal the Start tier's entire monthly spend cap. For a week of building, that is ample; for production, it is a trial. Watch Cerebral Valley's listings and Anthropic's events page, apply with a specific project rather than a general interest, and register with the Console organization your company will keep building in, so that whatever you build and win stays in one place.

7. Route 5: research, education, nonprofit and open-source programs

Anthropic runs several credit programs that are not aimed at startups but matter when a founding team includes researchers, maintainers or a nonprofit arm. The largest is AI for Science: "If successful, we will apply up to $20,000 in API credits to your account for a 6-month period" - Claude Help Center. It is designed for "researchers attached to research institutions working on high-impact scientific projects, with a particular focus on biology and life sciences applications", Anthropic evaluates submissions on the first Monday of each month, and the credits apply to the API, not the Claude web app.

Thematic calls inside AI for Science can include companies. Anthropic's rare-disease call offered "up to $50,000 in Claude credits over six months" and had "two tracks: one for scientists doing basic research, and another for early-stage biotechs working on speeding up clinical development for rare diseases" - Anthropic. Applications closed on August 2, 2026, but the existence of a startup track is the signal: an early-stage biotech should watch for the next thematic call rather than assume research credits are only for universities.

Two smaller research programs complete the API-credit side. The External Researcher Access Program applies "$1000 in API credits" for AI safety and alignment research, reviewed monthly - Claude Help Center. The Economic Futures Research Awards pair grants of $10,000 to $50,000 with "$5,000 in Claude API credits for computational analysis", though the page currently says "We are not currently accepting applications" - Anthropic.

The remaining programs discount subscriptions and seats, not the API. They are worth knowing because founders regularly mistake them for credit programs:

The open-source program is the one founders most often qualify for without realizing it. A founder who maintains a widely used library can get six months of Claude Max 20x personally, and a Max plan includes Claude Code - Claude. The program's terms limit it to natural persons, not companies, cap it at 10,000 recipients and charge overages at standard rates - Anthropic. For a startup, the practical value is one engineer's coding agent for half a year, not the product's API bill, and that is the right way to count it when planning a budget.

8. Route 6: AWS Activate and Claude on Amazon Bedrock

AWS is the one large cloud whose startup credits pay for Claude, and it says so in plain words. The AWS Startups FAQ states that "AWS Activate Credits are also redeemable for third-party models on Amazon Bedrock" and names Anthropic among the model providers - AWS Startups FAQ. The Activate terms, last updated January 22, 2026, make the mechanism precise: credits "may be applied to offset fees and charges for AWS Marketplace incurred for the use of third-party foundation models available on Amazon Bedrock", and "AWS Marketplace is an Eligible Service solely when incurring Bedrock 3P Model Spend" - AWS Activate Terms.

The word "solely" carries the whole rule. Claude on Bedrock is "a third-party model offered and billed through AWS Marketplace", and its charges "appear on your AWS bill and in AWS Cost Explorer under the model provider (not under Amazon Bedrock)" - AWS Bedrock documentation. Ordinary AWS promotional credits exclude AWS Marketplace unless AWS authorizes otherwise - AWS Promotional Credit Terms. The Activate terms are that authorization, for Bedrock's third-party models and nothing else. So Activate credits pay for Claude when you call it through Bedrock, and they do not stretch to other Marketplace purchases, including other ways of buying Claude on AWS.

How much, and who qualifies

AWS publishes three levels of credit on its credits page, and only the first two are open applications - AWS. The middle one, Portfolio, is where the large amounts are, and it is unlocked by an Organization ID from an "Activate Provider, such as an accelerator, angel investor, and venture capital firm", which makes it far more accessible than Anthropic's institutional-investor condition: an angel investor can be an Activate Provider.

The general conditions are "Pre-series B, founded in the last 10 years, AWS Account on Paid Tier Plan, and either new to Activate Credits or requesting more credits than previously received". Portfolio applicants "must apply within 12 months of your most recent funding date" - AWS, and the amount granted depends on the provider's offer. The credits page does not say whether the invite-only AI credits follow the same Bedrock rule, so confirm with your account manager before planning Claude spend against them. Our AWS credits guide walks through the Activate application step by step.

Using Claude on Bedrock with credits

Three setup steps stand between a credit balance and a Claude call. Anthropic models require a one-time First Time Use form before invocation, the account needs AWS Marketplace permissions, and "Your AWS account must have a valid payment method configured for AWS Marketplace purchases", even when credits will pay the charges - AWS Bedrock documentation. The account must also be on a paid plan, because "Free account plans are not eligible for Activate promotional credits". None of this is hard, but founders who apply for Activate, wait for approval and only then discover the form lose days they did not need to lose.

Prices on Bedrock match Anthropic's own for the global endpoints. AWS's Bedrock pricing lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, Claude Opus 5.5 at $4 and $20, and Claude Haiku 4.5 at $1 and $5 on global cross-region inference - Amazon Bedrock pricing. Regional endpoints cost more: "Regional and multi-region endpoints include a 10% premium over global endpoints" - Claude Docs, so Claude Sonnet 5.5 in US East runs $2.20 and $11, and a credit balance burns 10% faster there. Bedrock's batch inference is 50% cheaper than on-demand for select models, Anthropic's included.

Four rules from AWS's FAQ decide how much of an Activate grant actually reaches Claude. Credits "cannot be used to offset or cover historical AWS bills", so usage before the credits land is paid in cash. They "usually expire within 1-2 years depending on the package", and the Activate team "is unable to extend the expiration date". A larger second application pays only the difference: a startup that received $10,000 and later qualifies for $100,000 receives $90,000. And the terms say credits "are for your use only, and cannot be passed on to any end customers", which matters for agencies and studios that build Claude products for clients: the client's usage cannot run on your startup credits.

Claude Platform on AWS is a different product

Since May 11, 2026, AWS has offered a second way to buy Claude: Claude Platform on AWS, Anthropic's own platform, billed through AWS Marketplace - AWS Machine Learning Blog. Anthropic's documentation draws the line clearly: Claude Platform on AWS is "The Claude API platform billed through AWS Marketplace. Managed in the Claude Console and the AWS Console", while Claude in Amazon Bedrock is "An AWS-native service" - Claude Docs. AWS's billing page adds that usage is metered in Claude Consumption Units and that "CCUs are not prepaid credits; there is no CCU balance or commitment" - AWS.

No AWS or Anthropic page we read says Activate credits apply to Claude Platform on AWS, and the Activate terms' "solely when incurring Bedrock 3P Model Spend" points the other way. Anthropic's own pricing documentation lists the platform's payment model as "Arrears only (postpaid); no prepaid credits", so Anthropic's startup credits have no documented path there either. Until AWS says otherwise in writing, run credit-funded Claude usage through Bedrock.

Kiro: Claude inside AWS's coding agent

AWS offers one more indirect route. "AWS Activate Credits can also be used for Kiro Pro+", AWS's coding agent, and Kiro gives access to "frontier and open weight AI models from OpenAI, Anthropic, and other providers" - Kiro. Separately, Kiro Startup Credits give early-stage to Series A VC-backed startups up to a year of Kiro Pro+, but "Startups with active AWS Activate Credits are not eligible for this promotion" - Kiro. Either way, this is Claude inside a development tool, which helps your engineers, not API access that your product can call.

How to apply this section: if your product can call Claude through Bedrock, apply for Activate Founders now (it needs no investor), complete the First Time Use form and the Marketplace setup while you wait, and use the global endpoint unless a data-residency rule forces a regional one. When you raise from an investor who is an Activate Provider, apply for Portfolio within twelve months of the round. That sequence puts Claude on credits months before Anthropic's own program would consider you.

9. Route 7: Google Cloud, where startup credits stop at Gemini

Google's startup program is one of the most generous in the market, and it does not pay for Claude. Its early-stage page puts the rule in a footnote: "Program credits cover Google's state-of-the-art models like Gemini and Gemma. Note: Third-party models are billed directly and are not covered by the program credits" - Google Cloud. The program FAQ widens the exclusion to everything Google does not sell itself: the credits "cannot be applied to any third-party services or offerings including those on Google Cloud Marketplace" - Google Cloud FAQ.

Claude sits squarely inside that exclusion. Google's documentation names "Anthropic's Claude and Mistral models" as examples of "third-party managed models" - Google Cloud documentation, and its Claude page states that "Anthropic models are not a Google product" and are paid for "as you use them (pay as you go), or you pay a fixed fee when using provisioned throughput" - Google Cloud documentation. One naming change helps when reading older guides: Vertex AI is now "Gemini Enterprise Agent Platform (formerly Vertex AI)" - Google Cloud, which is why Anthropic's FAQ and Google's documentation seem to describe different products.

The one exception: a $10,000 partner-model perk

Google does carve out a small exception, and only in one sentence of one FAQ. "Qualifying startups who join the Scale and Scale AI Tier of the Google for Startups Cloud Program may receive $10,000 USD of credits that can be used toward partner models in Model Garden (including Anthropic, Mistral, AI21)", requested through your Google Cloud Account Executive - Google Cloud Startup Perks. Three caveats apply: the wording is "may receive", the credits must be requested rather than arriving with the program, and the perk grid on the same page showed no Anthropic card when we read it. Ask your Account Executive to confirm it in writing before you plan Claude spend against it.

For scale, the main program is large. The pre-funded Start tier gives "$2,000 in Cloud credits", valid for 12 months - Google Cloud. The funded Scale tier gives $200,000, and up to $350,000 for AI startups, released as up to $250,000 in the first year and a 20% credit on second-year spend worth up to $100,000 more. Of all of that, only the $10,000 perk can reach Claude. Claude's list prices on Google's global endpoint match Anthropic's own, $2 and $10 per million tokens for Claude Sonnet 5.5 and $4 and $20 for Claude Opus 5.5 - Google Cloud pricing, and program members pay them in cash.

How to apply this section: a startup on Google's credits can still build on Claude, it just should not expect Google to pay for it. Run Gemini or Gemma on credits wherever they do the job well, keep Claude for the tasks where it earns its price, and call Claude through the first-party Console, where Anthropic's own credits work, or through Bedrock if you also hold AWS credits. Our Google for Startups guide covers the program's tiers and application in depth, and our Google for Startups Cloud program page tracks its current terms.

10. Route 8: Microsoft Foundry, where Claude bills the card

Microsoft is even more explicit than Google. Its sponsorship policy reads: "Microsoft for Startups sponsorship credits apply to Microsoft Foundry models that are sold and billed directly by Azure. Models billed by third-party providers, partner services, or available through the Azure Marketplace are not eligible" - Microsoft Learn. The same page lists the providers whose models are not covered, and Anthropic appears without the "select models" qualifier some other providers carry: none of Claude is covered.

The detail that costs founders real money sits on the Claude deployment page. Among the subscription types "currently not supported" for Claude are subscriptions without an active pay-as-you-go billing method, such as student, free trial and startup credit-based accounts, and "Sponsored subscriptions that only use Azure credits", followed by a note: "If you have an account with a credit card on file, the credit card will be charged instead of Azure Credits" - Microsoft Learn. A startup that adds a card to its sponsored subscription and deploys Claude pays for every token with the card while its credit balance sits untouched.

The reason is, again, who sells the model. For both of Foundry's hosting options, "Anthropic is the seller and operator of Claude models in Microsoft Foundry. Claude models are Non-Microsoft Products under the Product Terms" - Microsoft Learn. Claude bills through Azure Marketplace in Claude Consumption Units, and that meter "decrements your Microsoft Azure Consumption Commitment (MACC)" - Microsoft Learn, which is a paid enterprise commitment, not a startup credit. The contrast case proves the rule: "Because Fireworks models are billed and sold through Azure, startup credits will apply to usage of models" - Microsoft Learn. Coverage follows the seller, not the model.

Microsoft's credits are still worth having for other work. The program unlocks up to $150,000 over time, and its Getting Started page lists milestones of $200, $5,000 after business verification, $25,000, $50,000 and $150,000, the higher ones tied to sustained Azure usage - Microsoft Learn; the program FAQ, updated the same day, describes the entry offer as $1,000 for 90 days and then $4,000, so two Microsoft pages disagree on the first step. Those credits cover OpenAI models that Azure sells directly, which our OpenAI API credits guide covers in detail, and our Microsoft for Startups program page tracks the current terms.

One indirect path remains. Azure startup credits can pay for usage-based GitHub services once you connect your Azure subscription as GitHub's payment method, and GitHub Copilot offers Claude models among others - GitHub Docs. No Microsoft or GitHub page we read states that Claude usage inside Copilot is specifically paid by startup credits, so treat it as plausible but unconfirmed, and in any case as a coding tool for your engineers, not API access for your product.

11. Route 9: gateways, platforms and banking perks

The last group of routes pays for Claude through a company that resells it. Gateways and developer platforms sit between your code and the model provider, and a credit on their bill pays for whichever model you route to, Claude included. These are the most portable credits in this guide, because one balance works across hundreds of models, and with one exception they are the smallest. Almost all of them are also gated by a referral from a partner, usually your bank, your accelerator or your investor.

That gate is the practical difference from the cloud programs. AWS, Google and Microsoft run open applications; gateway programs grow through partners who hand their perks to the startups they already serve. The founder's job is to find which of their existing providers is a partner, which takes one email rather than one application.

OpenRouter for Startups

OpenRouter's program is the best-known gateway route. "Eligible startups get up to $3,000 in universal inference credits usable across leading frontier and open models with 0% processing fees" - OpenRouter. To qualify you must be pre-Series B, build an AI-native product full-time, be "Referred by an approved OpenRouter partner", have a public website and a company-domain email, have "less than $500 in lifetime OpenRouter spend", and not have received startup credits before. The program terms, updated August 22, 2026, allow "credits in an amount up to $5,000 USD and/or waived platform fees for a period of up to twelve (12) months", say credits "may not be combined ("stacked") with other promotions unless explicitly allowed in writing", and expire unused credits at the end of their validity period - OpenRouter.

Claude costs the same on OpenRouter as at Anthropic: $2 and $10 per million tokens for Claude Sonnet 5.5 and $4 and $20 for Claude Opus 5.5 - OpenRouter. Mercury is one route to the referral: its perk offers "$1K in universal inference credits, valid for 6 months" with "$0 processing fees for 12 months", and notes that "Each applicant is allowed to receive startup credits once, even if they are affiliated with multiple partners" - Mercury. Secret, a deal marketplace, offers "$1,000 in credits across Anthropic's models through OpenRouter's API" to its Premium members, at $149 a year - Secret. Our OpenRouter for Startups page tracks the program, and our partner perks page lists Mercury's and Ramp's partner offers in full.

Vercel, Cloudflare, Databricks and Snowflake

Vercel's startup credit is the largest platform credit that pays for Claude. Vercel offers "an enterprise grade contract with Flexible Commitment Amount up to $30,000" usable on AI Gateway among other things, but "You may only use up to 50% of your Flexible Commitment Amount towards Vercel AI Gateway usage", for startups accepted on or after August 6, 2026 - Vercel. AI Gateway "charges no markup and no platform fee on tokens" - Vercel docs, so up to $15,000 of the award can buy Claude at list price. Access requires an approved Vercel Startups Partner, a round no later than Series A and an application within 12 months of it; see our Vercel for Startups page.

Cloudflare is the cautionary case. Its startup program grants credits in $10,000, $100,000 and $350,000 tiers, but its FAQ says "AI Gateway: Temporarily not covered by credits" - Cloudflare, and Workers AI, which the credits do cover up to a cap, does not offer Claude. Calling Anthropic through Cloudflare's Unified Billing means buying Cloudflare credits with "A 5% fee" on each purchase - Cloudflare docs. Our Cloudflare for Startups page covers what the credits do pay for.

The data platforms are the open question. Databricks says "Qualifying startups can receive up to $200K in credits for Databricks and Neon" - Databricks, and it serves Claude models through pay-per-token endpoints in your workspace - Databricks documentation. Snowflake offers eligible startups "free usage on their first contract" - Snowflake, and its Cortex REST API gives access to Anthropic's models - Snowflake documentation. Neither company says its startup credits cover Claude usage, so if your data already lives on one of them, ask before you plan Claude spend against the credits. Our Databricks and Snowflake pages track both programs.

Banking perks and deal marketplaces

One bank offers first-party Claude credits as a perk. Rho's perk page lists "$500 in Anthropic API credits to build with Claude", redeemable by signing in, and says perks "unlock the day your account opens" - Rho. It is the only first-party Claude API credit we found offered by a bank. Brex's perks directory listed no Anthropic or Claude offer when we read its pages on October 5 - Brex, and Ramp's rewards page offers OpenAI API credits but nothing from Anthropic - Ramp. Our business bank account comparison weighs each bank's perks against its fees and features, which is the right order: no bank is worth choosing for a $500 credit.

Deal marketplaces need more care. FounderPass lists an "Anthropic API discount" of "up to 20% or more off Anthropic API costs through our partner LLMAPI", a reseller rather than Anthropic - FounderPass. Its separate Claude credits page still describes "key Claude models, including Claude 3.5 Sonnet and Opus" - FounderPass Claude page, naming a model that appears in neither the current nor the legacy lineup on Anthropic's models overview. That does not make the offer worthless, but it shows how quickly deal pages drift. Before redeeming any marketplace offer, check who actually issues the credit, which platform it runs on, and whether it touches Anthropic at all.

How to apply this section: collect gateway and platform credits after the big routes, not instead of them. Their best use is experimentation across models and providers, where one balance on one bill saves real engineering time, and their partner gates mean you usually get them through a provider you already use. Start by asking your bank, your accelerator and your hosting platform whether they are partners of OpenRouter or Vercel.

12. What Claude credits buy: the burn math

A credit balance only means something once it is converted into work. Anthropic's current lineup has four models, and its models overview recommends a default: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads. Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short" - Claude Docs. Prices differ by a factor of ten from the cheapest to the most capable, so the model you route to matters more than almost any other cost decision.

The table below lists the current list prices per million tokens, from Anthropic's pricing documentation - Claude Docs. Cache reads are the price of re-reading context you have already cached, and batch prices apply to requests sent through the Batch API.

A few facts from Anthropic's launch pages and pricing documentation change how those prices compare. Claude Opus 5.5 costs "20% less than Opus 5", and its cache reads, "which make up the majority of agentic and coding work costs", are "$0.20 per million tokens, 60% less than Opus 5" - Anthropic. Claude Sonnet 5.5 is "priced the same as Sonnet 5" but "typically needs far fewer tokens to do the same work" - Anthropic. And newer models use a tokenizer that "produces approximately 30% more tokens for the same text", so per-token prices across model generations are not directly comparable. The 1M-token context window carries no surcharge: "A 900k-token request is billed at the same per-token rate as a 9k-token request" - Claude Docs.

A worked example: a customer support agent

Consider a support agent whose system prompt and tool definitions total 5,000 tokens, shared across every conversation, plus 1,000 tokens of conversation and a 400-token reply per turn. Without caching, every turn pays full input price for all 6,000 tokens. With caching, the shared 5,000 tokens are read from cache at the cache-read price, and steady traffic keeps refreshing the cache, so the one-time cost of writing it is negligible. On Claude Sonnet 5.5, an uncached turn costs $0.016 and a cached turn $0.007; on Claude Opus 5.5, $0.032 and $0.013.

Two levers dominate the result. Caching more than doubles the turns per dollar on every model, and the choice of model moves the result by a factor of ten: a cached Claude Haiku 4.5 agent answers about 285,700 turns per $1,000, while an uncached Claude Fable 5.1 agent answers 12,500, a gap of nearly 23 times. In credit terms, the $25,000 that comes with an Anthology Fund investment runs about 3.6 million cached Sonnet 5.5 support turns, and a $500 hackathon credit about 71,400. The example is deliberately simple (it ignores tool calls, retries and longer replies), so treat it as a way to reason about your own workload rather than a forecast of it.

A coding agent burns differently

Agentic and coding work has a different shape, and Anthropic's own launch page names the cost that dominates it: cache reads. A coding agent re-reads its growing context on every step. Take a step that re-reads 200,000 tokens of context, adds 2,000 new input tokens and writes 1,000 output tokens on Claude Opus 5.5. With caching, the step costs about $0.068 ($0.04 of cache reads, $0.008 of new input and $0.02 of output, ignoring the small premium for writing new context into the cache), or roughly $68 per thousand steps. Without caching, the same step costs about $0.83, or $828 per thousand steps, twelve times as much.

That ratio is why the cache-read price, not the headline input price, is the number to compare when choosing a model for agents, and why Anthropic cut it on both Claude Opus 5.5 and Claude Fable 5.1. It is also why a credit grant that looks modest can carry a coding-heavy startup a long way, as long as its engineers keep their agents' context cacheable. The arithmetic here is ours, from list prices; your own logs will show your real token mix within a day of usage.

13. Making credits last: caching, batching and model choice

Credits run out on a date, so stretching them is worth engineering time. Anthropic's own cost-optimization list starts with model choice: "Choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning" - Claude Docs. The models overview, as quoted in section 12, recommends starting with Claude Opus 5.5 for most workloads. The two pieces of advice stop conflicting once you count cost per task instead of cost per token: a more capable model often finishes in fewer steps, with fewer retries, and a cheaper model that needs three attempts costs more than an expensive one that needs one.

Anthropic's short video on choosing a model makes the same argument with Claude Fable 5.1 as the example: picking the cheapest model at the lowest effort looks like the obvious way to save money, and often is not. It is worth four minutes before you set a default model for your product.

How to apply this: pick two or three model and effort settings, run your own evaluation tasks through each, and compare the cost per successful task. Plan for change as well: Claude Haiku 5.5 "will join the Claude 5.5 family in the coming weeks" - Anthropic, Claude Haiku 4.5's retirement date is "Not sooner than October 15, 2026", and Claude Sonnet 4.5 was deprecated on September 30, 2026 with retirement set for November 30 - Claude Docs. A credit grant that outlives the model you built on is common, so keep model IDs in configuration rather than in code.

Prompt caching

Prompt caching is the largest single lever, and its prices are set out in one table. A five-minute cache write costs 1.25 times the base input price, a one-hour write twice the base price, and a cache read 0.1 times the base price, or 0.025 times on Claude Fable 5.1 and 0.05 times on Claude Opus 5.5 - Claude Docs. Anthropic spells out the break-even: "caching pays off after one cache read for the 5-minute duration (1.25x write), or after two cache reads for the 1-hour duration (2x write)." Any prompt prefix your application sends twice within five minutes should be cached.

What to cache follows from what repeats: the system prompt, tool definitions, long reference documents and the earlier turns of a conversation. Put the static parts first and the changing parts last, because the cache matches on the beginning of the prompt. Caching also lifts your rate limits, since cached input tokens do not count toward input-token limits on most models, as section 2 showed. Caching is also where a long-running bill is most often lost by accident, through small prompt changes that quietly stop the cache from hitting. A prompt caching walkthrough for Anthropic, OpenAI, Gemini and DeepSeek, which lists seven ways teams silently lose the cache, is published on Founden.ai, a site run by the same founder as StartupPerks.

Batch processing

The Batch API halves the price of anything that can wait. It "allows asynchronous processing of large volumes of requests with a 50% discount on both input and output tokens" - Claude Docs. A batch holds up to 100,000 requests or 256 MB, and batches expire if they do not finish within 24 hours - Claude Docs. Batch and caching discounts stack, so a nightly job that re-reads the same instructions for thousands of records pays half price on the uncached part and a fraction of that on the cached part. Evaluation runs, data enrichment, classification, content generation and back-office summaries are all natural batch work, and moving them off the synchronous API is usually a day of engineering for a permanent halving of their cost.

The premiums to avoid while on credits

Several options raise the per-token price, and each is worth using only on purpose. US-only inference through the inference_geo parameter "incurs a 1.1x multiplier on all token pricing categories" on recent models. Regional endpoints on Bedrock and Google Cloud carry the 10% premium described in section 8. Fast mode, a research preview, prices Claude Opus 5.5 at $8 per million input tokens and $40 per million output tokens, twice the standard rate, and is not available with the Batch API. Server tools carry their own prices: web search costs "$10 per 1,000 searches, plus standard token costs for search-generated content", web fetch carries no additional cost, and code execution includes 1,550 free hours per organization each month before billing at $0.05 per container-hour - Claude Docs.

None of these is wrong: data residency can be a legal requirement, and latency can be the product. The point is that each one burns a grant faster, so make each a deliberate decision with a reason written down, rather than a default someone set during a prototype and nobody revisited.

14. Which route fits your startup

Two questions decide almost everything in this guide. The first is whether you have institutional equity funding and are under four years old, because that is the condition on Anthropic's own credits. The second is whether your product can call Claude through Amazon Bedrock, because that decides whether the largest cloud credit program pays for your Claude usage. Accelerators, hackathons, gateways and perks are supplements around those two answers, not substitutes for them.

The diagram below turns those questions into a path. It is deliberately simple: a real company will often take two branches at once, for example applying to Anthropic's program and to AWS Activate in the same week, and every branch ends in the same place, because no route lasts if the credits are burned carelessly.

Most startups fit one of six profiles, and each has a clear first move.

The bootstrapped founder cannot get Anthropic's credits, because there is no institutional investor, but should still join the program for its community and events. The first real credit is AWS Activate Founders: $1,000 with no investor required, spent on Claude through Bedrock. Hackathons add $500 at a time, and Rho's $500 perk is there if you bank with Rho anyway. With prompt caching on Claude Sonnet 5.5, $1,000 covers about 142,900 turns of the support agent in section 12, which is enough to find out whether the product works.

The angel-funded pre-seed team is in an awkward spot with Anthropic, whose condition names an institutional investor, and a strong one with AWS. Activate Portfolio is unlocked by an Organization ID from an "Activate Provider, such as an accelerator, angel investor, and venture capital firm", so the first question for each angel is whether they are a provider. If one is, Portfolio can be worth far more than anything else on this list.

The seed-stage company with an institutional lead should apply to Claude for Startups and name the lead, ask the lead whether they are an Anthropic partner, and apply for Activate Portfolio within twelve months of the round. Those two grants come from different companies and pay different bills, so they stack, and together they usually cover the first year of model spend.

The Series A company already on AWS has the easiest path: Claude on Bedrock against Portfolio credits, Claude for Startups if the company is still within four years of founding, and a conversation with its AWS account manager about the invite-only AI credits once spend justifies it. Its main risk is the Scale tier's spend cap on the first-party API and the 10% regional premium on Bedrock, both covered above.

The startup holding Google or Azure credits should not move Claude onto those bills. Run Gemini, Gemma or Azure-sold OpenAI models on the credits where they fit, keep Claude on the first-party Console or Bedrock, and, on Google's Scale tiers, ask the Account Executive about the $10,000 partner-model perk.

The academic spinout or research team should look at AI for Science first: up to $20,000 over six months for qualifying researchers, plus thematic calls that have included a track for early-stage biotechs. It is the largest Anthropic credit with a published amount that does not require a venture investor.

15. Stacking routes, and the order to apply

Most of these routes stack, because they come from different companies and pay different invoices. Anthropic's credit pays your Console bill, AWS's pays your Bedrock line, OpenRouter's pays your OpenRouter balance. Nothing in Anthropic's terms forbids holding AWS credits, and nothing in AWS's terms forbids holding Anthropic's. A well-run startup can therefore run Claude on several balances at once and route each workload to whichever balance has room.

Four rules limit the stacking, and each has caught founders out. Anthropic's grant is once per company ("not previously received Anthropic startup credits"). OpenRouter's credits come once per applicant even across partners, and do not combine with other OpenRouter promotions unless OpenRouter agrees in writing. AWS pays only the difference on a larger second Activate application, and Kiro's startup credits are closed to anyone with active Activate credits. And no credit pays retroactively: AWS says outright that Activate credits cannot cover historical bills, so spend before approval is paid in cash. Those rules, and the conditions on each program, give a natural order.

The first move costs nothing: set up the Console organization you will build in, put real Claude traffic through it and note its API Organization UUID. Anthropic weighs Claude integration and usage, and an organization with a few weeks of production history is the best evidence an application can carry. In the same week, apply for AWS Activate Founders, which needs no investor, and complete Bedrock's First Time Use form and Marketplace setup while the application is reviewed, because Activate credits take time to land and never pay for usage that came before them.

The funded applications wait for the round, because both programs are conditioned on it. After an institutional round, apply to Claude for Startups and name the lead investor, then apply to Activate Portfolio within twelve months of the round, using the Organization ID from whichever investor or accelerator is an Activate Provider. Applying to Anthropic earlier spends its one-time grant on a company that cannot yet use it well, and applying to AWS later than twelve months after the round forfeits the Portfolio track for that round.

Gateway credits and hackathons fill the gaps. Add a gateway credit through your bank or accelerator if you route work across many models, where one balance across providers saves real engineering time, and enter hackathons and watch for cohorts such as the Vercel and Agentic AI accelerators, which have offered equity-free Anthropic credit for a limited group on a fixed timeline. None of these needs to be planned months ahead; each is worth a short application when it appears.

Here is how the sequence plays out for a typical seed-stage company. It opens its Console organization in month one and ships a prototype on Claude Sonnet 5.5, takes $1,000 from Activate Founders in month two and moves batch work to Bedrock, and raises a seed round led by an institutional fund in month five. In month six it applies to Claude for Startups naming the lead, and to Activate Portfolio with the lead's Organization ID. Each grant pays a different invoice, so none of them cancels another, and the company reaches its Series A with most of its model spend paid by two companies' credits rather than its own cash.

Build your code so the credits can follow the workload. The same model has different IDs on different platforms: Claude Opus 5.5 is claude-opus-5-5 on the first-party API and anthropic.claude-opus-5-5 on Amazon Bedrock, according to the ID rows of Anthropic's models overview - Claude Docs. Keep model calls behind one internal client with the platform and model ID in configuration, so that moving a workload from the Console to Bedrock when one balance runs low is a configuration change rather than a refactor. Prices for the current models match across the first-party API, Bedrock's global endpoints and Google's, so the routing decision is about which balance has credit, not which platform is cheaper.

16. Eight mistakes that waste Claude credits

Most wasted Claude credit is not spent on bad ideas; it is spent on the wrong bill. The pattern in every mistake below is the same: a founder assumes credits follow the model, when they follow the invoice. Each of these happened often enough that the owners' own documentation warns against it, usually in a footnote.

The second pattern is timing. Credits expire, caps reset monthly, and approvals take time, so a grant can be lost without a single wasted token. Read the eight mistakes as a pre-flight check before you start spending.

1. Expecting Google or Microsoft startup credits to pay for Claude. Both exclude third-party models in writing. Google's $10,000 partner-model perk is the only exception, and it must be requested through an Account Executive.

2. Deploying Claude on a sponsored Azure subscription with a card on file. Microsoft's Claude page says the card "will be charged instead of Azure Credits". The bill arrives on the card, and the credit balance looks untouched, so the mistake can run for weeks before anyone notices.

3. Buying Claude Platform on AWS with Activate credits in mind. The Activate terms cover AWS Marketplace "solely" for Bedrock third-party model spend, and no page says the Anthropic-operated platform qualifies. Use Bedrock for credit-funded Claude.

4. Leaving a regional endpoint on by default. Regional and multi-region endpoints on Bedrock and Google Cloud cost 10% more than global ones. Unless a data-residency rule requires it, the global endpoint makes the same credit last longer.

5. Running without prompt caching. In the support-agent example in section 12, caching more than doubles the work per dollar on every model, and in the coding-agent example it cuts the cost twelvefold. It is usually a few lines of code.

6. Sending every request to the most capable model. Claude Fable 5.1 costs ten times as much per token as Claude Haiku 4.5. Measure cost per successful task on your own evaluations and route simple work to cheaper models.

7. Ignoring the Start tier's $500 monthly spend cap. Credits do not raise the cap. A new organization with a large grant can still only spend $500 a month until its tier rises or Anthropic upgrades its limits, so request higher limits as soon as real usage justifies them.

8. Letting credits expire, or spending before they land. Anthropic's promotional credits expire on the date stated at issue, or after a year; Activate credits usually last one to two years with no extensions; and neither pays for usage before it was granted. Put the expiry dates in your calendar the day the credits arrive.

The first three mistakes are about where Claude runs, the next three about how it runs, and the last two about when you spend. A fifteen-minute review of those three questions before the first production deploy prevents nearly all of them. Our startup credits playbook applies the same review to every credit a startup holds, and our cloud credits comparison shows how the three clouds' programs compare beyond Claude.

17. Frequently asked questions

How much does the Anthropic startup program give? Anthropic does not publish an amount. The program page offers "free credits and priority rate limits" to venture-backed startups without a figure, and the only startup credit amount Anthropic states on its own site is the $25,000 that comes with an investment from the Menlo Anthology Fund. Figures of $5,000, $25,000 or $100,000 tiers come from third-party sites.

Do I need VC funding to get Claude credits? To join the program, no: any early-stage founder can join. To receive Anthropic's startup credits, yes: the company must have "received equity funding from institutional investor", have been founded within the last four years, and not have received Anthropic startup credits before. Bootstrapped founders can still get credit that pays for Claude through AWS Activate Founders, hackathons and some perks.

Can I use Anthropic's startup credits on AWS Bedrock or Google Cloud? No. Anthropic's FAQ says they "apply to the first-party Claude API via the Claude Console" and "can't be used on AWS Bedrock, Google Cloud Vertex AI, or other third-party platforms."

Do AWS credits pay for Claude? Yes, on Amazon Bedrock. AWS Activate credits are "redeemable for third-party models on Amazon Bedrock", and AWS names Anthropic among the providers. They do not have a documented path to Claude Platform on AWS.

Do Google Cloud or Microsoft for Startups credits pay for Claude? No. Google's program credits exclude third-party models, with a separate $10,000 partner-model perk available on request to Scale and Scale AI tier members. Microsoft's sponsorship credits cover only models sold and billed directly by Azure, and Microsoft lists Anthropic among the providers not covered.

How long do Claude startup credits last? The program page does not say. Under Anthropic's Supplemental Credit Terms, promotional credits expire on the date indicated when they are issued, or one calendar year after issue if no date is given. Purchased credits expire one year after purchase.

Do Claude credits cover Claude Code? Yes, when you use Claude Code with a Console account, which bills API tokens at standard prices. Anthropic's billing article says prepaid credits "cover API access, playground usage, and Claude Code".

Does a Claude Pro or Max subscription include API credits? No. A paid Claude subscription "doesn't include access to the Claude API or Console", and the chat account and the Console account operate independently even on the same email address.

Do new API accounts get free credits? A small amount. Anthropic says "New users receive a small amount of free credits to test the API" without stating how much.

18. How we verified this guide

Every figure in this guide was read on its owner's own page on October 5, 2026. We fetched each program page directly, checked every quotation against the page text, and kept a figure only when the company that issues the credit states it: a number that appears only on a third-party deal site or blog was left out, or described as unverified. Model names, prices, rate limits and spend caps come from Anthropic's models overview, pricing and rate-limit documentation, read the same day. No application was submitted and no account was created to research this guide.

Several things could not be verified, and the guide says so where they matter: the size of Anthropic's startup grants, the list of Anthropic partner funds, the program's review time, the expiry of startup credits beyond Anthropic's general credit terms, whether Activate credits apply to Claude Platform on AWS, Google's $10,000 partner-model perk beyond one FAQ sentence, and the first step of Microsoft's credit ladder, on which two Microsoft pages disagree. While researching, we corrected our own catalog: our Claude for Startups page said "up to $100K" and now says the amount is not published.

StartupPerks is not endorsed by Anthropic, Amazon Web Services, Google, Microsoft or any other company named in this guide. No program paid to appear here, and this guide contains no referral links. For the wider picture of AI and model credits beyond Claude, see our AI credits for startups guide.

This guide reflects the programs, terms and prices on their owners' pages as of October 5, 2026. Credit programs change without notice, so confirm the current terms on each linked page before you apply.