AI credits for startups in 2026: what's actually on offer, and what isn't

Every guide to AI startup credits lists the same numbers. Most of them are wrong. Program pages get taken down, terms change, and aggregator sites republish the old figures indefinitely. This guide was built from vendor pages only: the amounts, gates, and caveats below come from the source or are marked "not published."

The central finding worth knowing upfront: Anthropic's credit program accepts public applications and its FAQ states VC funding is not required. OpenAI's program directs founders to their VC. Most guides have this backwards.

Last verified: August 2026. Program terms change frequently. Confirm details at each provider's site before making commitments.

AI model and API providers

OpenAI

What you get: Credit allocation for API usage. Dollar amount not published.

The gate: The program FAQ names its VC partners: Thrive Capital, Sequoia, a16z, Kleiner Perkins, and Conviction Partners. Portfolio companies of these firms get access to credits, rate limit upgrades, and time with solutions engineers. The FAQ instructs founders to reach out to their VC to unlock credits. If you are bootstrapped or pre-institutional, OpenAI's primary credit program is not accessible.

What is open: Build Hours, developer resources, starter repos, guides, and founder events are available to any startup without a VC referral. OpenAI's own FAQ says "You won't need a pitch deck," which contrasts with NVIDIA Inception requiring one.

OpenAI Grove: A separate five-week in-person program in San Francisco for pre-idea to early-stage founders. Approximately 15 participants per cohort. Participants receive $50,000 in API credits, early access to tools, and mentorship from OpenAI's technical team. Cohort 2 applications closed 12 January 2026.

Expiry: Per OpenAI's Service Credit Terms, credits "expire one year after the date of purchase or issuance if not used, unless otherwise specified at the time of purchase."

Source: openai.com/business/why-openai/startups/

Anthropic

What you get: API credits and priority rate limits for building on Claude. Amount not published on the program page.

The gate: Public application at claude.com/programs/startups. The FAQ explicitly states the program is open "with or without VC backing." Founders backed by an Anthropic partner VC or accelerator can mention it in their application and may qualify for additional benefits, but VC backing is not a requirement.

Eligibility that Anthropic does state: institutional equity funding in practice, founded within the last four years, and no prior Anthropic startup credits. You need a Claude Console account and a company email address.

The catch: Credits work only on the first-party Claude API via Claude Console. They cannot be used on AWS Bedrock or Google Vertex AI. This catches out teams planning to consolidate onto a hyperscaler. Accepted startups also get Anthropic's highest rate limits.

Source: claude.com/programs/startups - application form at claude.com/form/startups-application

Mistral AI

What you get: A consumer Free plan at mistral.ai/pricing that includes $10/month in API credits. The La Plateforme "Experiment" API tier has no published limits.

The gate: None for the Free plan. A startup program called "Mistralship" existed, but the application page returns 404 as of August 2026. No active public application process exists.

The catch: The €30,000 figure circulating across aggregator sites cannot be confirmed on mistral.ai. Do not treat it as current. The $10/month Free plan credit is the only confirmed offer.

Source: mistral.ai

Together AI

What you get: API credits through the Together AI Startup Accelerator: up to $15,000 if you've raised under $5M, $30,000 at $5–10M, $50,000 above $10M. Plus engineering resources and GTM support.

The gate: Funding-gated at every tier, which excludes bootstrapped companies. The program is for AI-native startups from seed through Series B building on open-source models. The tiers are gated on amount raised, not on a VC referral - a company that raised $500K from angels qualifies for the Build tier.

The catch: No free tier exists. Together AI's docs state there are no free trials and API access requires a minimum $5 purchase upfront. A zero balance suspends access. Expiry is not published.

Source: together.ai/startup-accelerator

Groq

What you get: A genuinely free tier with no credit card required and no credits system. Rate limits are the only constraint.

Free plan limits (confirmed from console.groq.com/docs/rate-limits, August 2026):

The "14,400 RPD across all models" figure that appears on aggregator sites is the llama-prompt-guard-2 limit, applied erroneously to everything else.

The gate: A startup credit program page existed and now redirects to Groq's homepage, which explains why stale figures circulate. Adding a credit card unlocks the Developer plan with roughly 10x higher limits.

The catch: Free-tier limits are sufficient for prototyping; they are not designed for production workloads. Groq's own Service Credit Terms specify a one-year expiry for promotional credits issued at events like hackathons.

Source: console.groq.com/docs/rate-limits (verified August 2026)

Fireworks AI

What you get: Startup program with credits for inference and training. Amount not published on the program page.

The gate: Registered for-profit company with a functional website, founded within the last five years, VC-backed with $500,000 or more raised, and not yet at Series D or later. Credits expire after one year.

The catch: $1 in signup credits is available to any new account. The startup program requires meeting the funding threshold, so it is not accessible to bootstrapped or pre-seed companies.

Source: fireworks.ai/startups

Google AI Studio / Gemini API

What you get: Free access to every Gemini model via AI Studio. The free tier lists every model as "Free of charge."

The gate: None for basic access.

The catch: Two things distinguish this from a simple free tier. First, figures of 15 RPM / 1,500 RPD still circulate widely but Google no longer confirms them - current limits are accessible after login. Second, content submitted on the free tier is used to improve Google's products; content on paid tiers is not. This is relevant for anything proprietary or sensitive.

This is separate from Google for Startups Cloud Program credits, which are hyperscaler infrastructure credits covered below.

Source: ai.google.dev

Cohere

What you get: A free Trial API tier at 1,000 calls per month, 20 requests per minute for chat models.

The gate: Trial keys are barred from commercial use. Moving to production requires a paid plan.

The catch: The startup program page on cohere.com now resolves to the homepage. The offer, which was discounts rather than credits, appears to have been withdrawn. Do not reference a Cohere startup credit program.

Source: cohere.com

Hugging Face

What you get: A free monthly Inference API allowance of $0.10/month on the free plan, $2.00/month on PRO, and $2.00 per seat on Team and Enterprise plans. Credits apply to HF-routed requests only.

The gate: No gate for the free tier.

The catch: $0.10/month is enough for very light experimentation. No startup program exists. The common aggregator framing of "free Inference API with daily rate limits" misdescribes the billing model - it is a monthly credit allowance, not a rate limit.

Source: huggingface.co

OpenRouter

What you get: Free access to model variants tagged `:free` across the catalog. An auto-router selects from available free models based on context. Purchased credits unlock access to paid models.

The gate: No program. Free model access is available to any registered user.

The catch: OpenRouter's own pages contradict each other on daily limits: 20 RPM / 200 RPD on the limits page versus 50 RPD without credits and 1,000 RPD with 10 or more credits in the FAQ. Treat the daily ceiling as unsettled. Purchased credits carry a 5.5% fee and a one-year expiry. No startup program exists.

Source: openrouter.ai

GPU and inference compute

RunPod

What you get: Two distinct tiers.

Growth tier: Deposit $50,000, receive a $25,000 bonus, giving $75,000 in total compute. One-time, 12-month agreement, no reload.

Starter tier: $1,000 in credits, no deposit required, no funding requirement.

The gate: RunPod's FAQ says venture backing is "strongly preferred" for the Growth tier, though they "review every application holistically." The Starter tier has no funding gate. The headline "$25K in credits" that circulates widely misrepresents the Growth tier: the bonus is $25K, but accessing it costs $50K upfront.

Source: runpod.io

Modal

What you get: A recurring $30/month in free compute on the free plan, confirmed on modal.com/pricing. The startup program exists and offers "thousands of free GPU credits," but the specific amount and expiry term are not published.

The academic program, which is separate, publishes up to $10,000.

The gate: Startup program requires either a partner-network VC or $1M or more raised.

Source: modal.com/pricing

Not published: Credit amount and expiry for the startup program.

Lambda Labs

What you get: $7,500 in cloud credits.

The gate: Restricted to NVIDIA Inception or NVIDIA Connect members. One grant per company. Geo-restricted. There is no free tier or trial credit separate from this program. A separate open research grant offers up to $5,000 with different eligibility criteria.

The catch: "$7,500 for any startup" is not accurate. Being an NVIDIA program member is a meaningful pre-qualification step. Apply to NVIDIA Inception first; that makes you eligible for the Lambda credit.

Source: lambda.ai

Baseten

What you get: Up to $25,000 for dedicated inference or training, plus up to $2,500 for Model APIs.

The gate: AI-first company, Seed through Series A, founded fewer than five years ago, net-new accounts only.

The catch: Expiry is not published. "AI-first" means AI is core to the product, not a feature added on top of something else.

Source: baseten.co/startup-program

Replicate

What you get: Free runs exist but Replicate does not publish the allowance. There is no confirmed startup credit program: replicate.com/startups does not resolve.

Source: replicate.com

Speech, voice, and media AI

ElevenLabs

What you get: 33,000,000 credits plus 12 months of Scale tier access. ElevenLabs values this at approximately $4,000+.

The gate: Fewer than 25 employees. No VC requirement. One of the most accessible gates among voice AI providers.

The catch: The free tier is 10,000 credits per month and is non-commercial only. An older page on elevenlabs.io has been superseded by the startup-grants page.

Source: elevenlabs.io/startup-grants

Deepgram

What you get: $200 in API credits on signup, no credit card required. The startup program offers up to $100,000 over 12 months.

The gate: No VC required for either the signup credit or the startup program. The program is for early-stage AI builders in production or launching within six months. The higher tier is issued "at Deepgram's sole discretion (including amount and timing)" and is usable only after standard credits are exhausted. Credits activate on Pay As You Go plans only. Deepgram's pricing page and its docs disagree on whether the $200 signup credit expires; the docs specify a one-year rule, which governs.

Source: deepgram.com/startup-program

AssemblyAI

What you get: 12 months of recurring credits. Amount not published on the program page. $50 in free credits with no card required; these do not expire.

The gate: Bootstrapped through Series A, no VC required. Founders in partner VC portfolios are fast-tracked, not exclusively eligible. Application at assemblyai.com/contact/startup-program, with a 2–3 business day review.

The catch: A legacy page advertising $5,000 and 35% off is still live at a different URL on assemblyai.com. Those terms are superseded. The current offer has an unpublished credit amount.

Source: assemblyai.com/contact/startup-program

Runway

What you get: 500,000 API credits for qualifying startups. The Runway Fund is a separate equity vehicle offering up to $500,000 in pre-seed or seed investment. It is not part of the credits program and should not be described as though credits and investment are bundled.

The gate: Credits: public application. Fund: equity investment in pre-seed or seed-stage companies, launched March 2026, selective.

The catch: API credits cannot be spent in the Runway web app - they are API-only. The free plan provides 125 one-time credits that are watermarked. These two pools are completely separate.

Source: runway.com

Hyperscaler AI credits

These are infrastructure credits, not model API credits. They cover compute, storage, and managed AI services rather than per-token API access.

For full detail on each, see our guide to startup credit programs compared.

Free tiers worth knowing

If you are pre-funding and program gates are a barrier, these free tiers offer real access with no credit card:

Groq: No card required, no credits system, all models accessible within rate limits. Best for prototyping with fast inference. Limits vary by model: llama-prompt-guard-2 gets 14,400 RPD; most production models cap at 1,000 RPD.

Google AI Studio: Every Gemini model listed as free. Figures of 15 RPM / 1,500 RPD still circulate widely but Google no longer confirms them - check ai.google.dev after login for current figures. Note the data-use clause: free-tier inputs are used to improve Google's products.

AssemblyAI: $50 in credits, no card, no expiry. The most durable free starting point among the speech providers.

Deepgram: $200 on signup, no card. Best for testing speech-to-text before committing.

Hugging Face: $0.10/month in Inference API credits. Useful for very low-volume experimentation on HF-hosted models.

OpenRouter: Free model variants available with no card. Limits are contradicted across OpenRouter's own pages; treat the daily ceiling as unsettled.

For more tools with genuinely free tiers, see our companion guide.

Stacking programs

Several combinations work well together and target different spending categories:

Model API layer: Anthropic startup program (no VC required, public application) or Together AI Accelerator (up to $50K, funding-gated) for primary model spend.

Compute layer: Modal's $30/month recurring free tier for low-volume jobs; Baseten startup program (up to $25K) for dedicated inference at scale.

Voice/audio layer: ElevenLabs startup grant (33M credits, no VC required) for synthesis; Deepgram $200 signup credit or AssemblyAI $50 no-expiry credit for transcription.

Infrastructure: Google Cloud ($350K AI-first), AWS Activate, or Microsoft for Startups for storage, managed services, and compute outside inference. Eligibility and stacking rules for all three are in startup credit programs compared.

Note on NVIDIA Inception: Free to join, pitch deck required, and it unlocks downstream programs including the Lambda Labs cloud credit. Exclusions include consulting firms, crypto companies, cloud providers, resellers, and public companies. No credit amounts are published directly by NVIDIA Inception: benefits come through program partners.

Figures you'll see elsewhere that we couldn't verify

These numbers appear on widely-read aggregator sites. None of them were confirmed on a vendor page in August 2026.

What AI credits don't cover

Credits buy inference, GPU compute, and API calls. They do not buy the application layer.

Helpdesk software, analytics and monitoring, error tracking, CRM, payments processing, and compliance tooling sit outside every program in this guide. These are the costs that persist after your credits expire and your prototype becomes a product.

Reserved GPU capacity is also commonly excluded. Together AI and Fireworks AI both exclude reserved clusters from startup credit programs. If your workload requires guaranteed capacity rather than on-demand inference, credits will not cover it.

This gap is worth planning for early. A startup that spends its first year optimizing model API costs while paying full price for everything around it will find the non-credit costs are often the larger line item by the time it reaches Series A.

Where the Fin Startup Pack fits

The Fin Startup Pack covers the application layer that credits leave out: helpdesk, AI agent, and the platform your support and sales teams run on. It bundles 55+ deals worth over $500,000 through a single application.

The gate is headcount alone: under 50 employees. No funding round requirement, no investor introduction, no accelerator membership. Pre-launch qualifies with a working prototype. The pack deliberately excludes hyperscaler compute, so apply to Google Cloud, AWS, or Microsoft directly for that.

Apply for the Fin Startup Pack

FAQs

What are AI credits for startups?

AI credits are prepaid allocations of API usage, inference compute, or platform access that providers offer to early-stage companies at no cost or reduced cost. They let startups build and test on production-grade AI infrastructure without paying full price during the earliest stages of development. Most programs have eligibility requirements around funding stage, company age, or headcount. Credits typically expire after 12 months and cover usage-based costs like API calls, GPU hours, or character generation. They do not cover seat-based software, compliance tooling, or the application layer that sits between your AI models and your customers.

Do I need VC backing to access AI startup credits?

It depends on the provider. Anthropic, ElevenLabs, AssemblyAI, and Deepgram all explicitly do not require VC backing. Together AI and Fireworks AI are funding-gated. OpenAI's primary credit program routes through VC partners. Groq's free tier requires nothing. RunPod's Starter tier has no funding requirement, but the Growth tier's $50K deposit effectively gates it to funded companies. For non-AI tools on the same terms, see startup deals with no VC required.

What is the difference between AI credits and cloud credits?

AI credits cover model API calls and inference compute: the cost of running prompts through language models, generating speech, or processing images. Cloud credits cover general infrastructure: virtual machines, storage, databases, networking, and managed services. Google AI Studio credits and Google Cloud for Startups credits are separate programs with separate applications. A startup that needs both model access and hosting infrastructure should apply to both.

What do AI credits not cover?

Credits buy inference, GPU compute, and API calls. They do not buy helpdesk software, analytics, monitoring, error tracking, CRM, payments processing, or compliance tooling. Reserved GPU capacity is also commonly excluded: Together AI and Fireworks AI both exclude reserved clusters. These are the costs that persist after credits expire.

Do AI credits require a credit card?

Groq, Deepgram, and AssemblyAI require no credit card for their free tiers or signup credits. Google AI Studio requires a Google account only. Modal and Baseten require a credit card for their free tiers. Together AI requires a $5 minimum purchase. OpenRouter offers free model access without a card but requires one for purchased credits.

Do AI credits expire?

Most do. OpenAI's Service Credit Terms specify one year from issuance. Fireworks AI specifies one year from acceptance. Deepgram's docs specify one year for signup credits. OpenRouter purchased credits carry a one-year term. Groq's Service Credit Terms specify one year for promotional credits issued at hackathons. AssemblyAI's $50 free credit does not expire. Modal, Baseten, and several startup programs do not publish expiry terms.

Can I stack credits from multiple providers?

Yes, for different spending categories. You might use Anthropic credits for model API calls, Modal or Baseten for dedicated compute, ElevenLabs for voice synthesis, and a hyperscaler for storage and managed services. Stacking credits from two programs at the same provider is subject to each provider's terms.

Is the Anthropic startup program genuinely open to bootstrapped founders?

Anthropic's FAQ states the program is open "with or without VC backing," and there is a public application route at claude.com/programs/startups. In practice, credit acceptance requires institutional equity funding, so a fully bootstrapped company may be admitted to the program without receiving credits. Partner-portfolio founders can note their investor and may qualify for additional benefits. Compared with OpenAI, which routes credits exclusively through partner VCs, Anthropic's route is the more open of the two.

Why do aggregator sites list different numbers?

Most aggregate from a snapshot taken when a program launched, then republish without updating. Program pages change or go offline; aggregators do not. The Mistral €30,000 figure, for example, is almost certainly from the original Mistralship launch announcement. The page 404s now.

What does "at Deepgram's sole discretion" mean for the $100K program?

It means the amount, timing, and even whether credits are issued at all are decided by Deepgram on a case-by-case basis. The $100K is a ceiling, not a guarantee. The $200 signup credit is unconditional.

Does the Groq free tier work for production workloads?

For most models, no. Most production models cap at 1,000 requests per day on the free plan. Adding a credit card and upgrading to Developer unlocks roughly 10x the limits. Llama prompt-guard models have a higher free-tier ceiling (14,400 RPD) because they are lighter classification models, not generation models.

What is NVIDIA Inception and why does it matter for GPU credits?

NVIDIA Inception is a free-to-join accelerator program for AI startups. A pitch deck is required to apply. Joining unlocks downstream benefits from NVIDIA partners including the Lambda Labs $7,500 cloud credit. NVIDIA does not publish credit amounts directly: you apply to Inception and then access partner benefits through the program. Consulting firms, crypto companies, cloud providers, resellers, and public companies are excluded.

Where should I look for the most current terms?

Always the provider's own program page. Check claude.com/programs/startups, openai.com/business/why-openai/startups/, together.ai/startup-accelerator, fireworks.ai/startups, baseten.co/startup-program, elevenlabs.io/startup-grants, deepgram.com/startup-program, assemblyai.com/contact/startup-program, and runway.com directly.

How does the Fin Startup Pack relate to AI credits?

AI credits cover the model and compute layer. The Fin Startup Pack covers the application layer on top: helpdesk, AI agent, and the operational tooling your team runs on. The gate is under 50 employees with no VC or accelerator requirement. Apply to Google Cloud, AWS, or Microsoft separately for infrastructure.