Free AI compute, found and checked. Who qualifies and the catch, up front.
API credits, free inference tiers, startup programs and research grants from the frontier labs and the long tail. Every offer is checked against the provider’s own page: value, eligibility, limits, deadline, fine print.
VC-backed startups building on Gemini can get up to $350,000 in Google Cloud credits over two years.
For: Startups that raised Pre-Seed/Seed within 5 years (or Series A within 12 months), founded within 5 years, built on Gemini, with under $5,000 in prior Google Cloud credits
The Gemini API has a free tier with free input and output tokens on Gemini 3.x Flash and Flash-Lite models, no card needed.
Anyone with a Google account
Per-project RPM/TPM/RPD, see AI Studio
Catch: Free-tier content may be used to improve Google products; paid tier does not. Rate limits are per project, can change and are not guaranteed. Flagship Gemini 4 Argon is paid only. Upgrading to Tier 1 requires linking billing.
Modal's Starter plan includes $30 of free compute credit every month for serverless GPUs, CPUs and sandboxes.
Anyone with a Modal account
Resets monthly
Catch: Usage beyond the $30 is pay-as-you-go. As of Sept 1, 2026 Shared Endpoint tokens are billed from the first request and are not covered by the included credit. Cloud credits from AWS/GCP/Azure cannot be applied to Modal.
Groq's free plan gives rate-limited API access, e.g. gpt-oss-120b at 30 requests/min and 1,000 requests/day.
Anyone with a Groq account
30 RPM - 1,000 req/day - 200K tokens/day
Catch: Limits are per organization and per model; the docs say exact limits are on your account's Limits page and there may be exceptions. Developer plan needs a payment method. Not suited to production. The docs table is labelled as base limits, so confirm yours.
Workers AI includes 10,000 Neurons a day at no charge on the free Workers plan, enough for light inference on open models.
Anyone with a Cloudflare account
10,000 Neurons/day
Catch: Free plan stops at the daily allocation; going over requires the Workers Paid plan. How far 10K Neurons goes depends on the model (large models burn it fast).
Mistral's free mode lets you create API keys and use a limited included monthly usage, no prepaid credits needed.
Anyone with a Mistral account
Catch: The docs don't publish the included usage or exact rate limits; they are shown in the Admin Panel. Verification steps (like a phone number) are not described in the docs. Upgrade and pay-as-you-go uses prepaid credits.
Cohere trial (evaluation) keys are free but limited to 1,000 API calls a month.
Anyone with a Cohere account
1,000 calls/mo - 20 chat req/min
Catch: Trial keys are meant for evaluation; production needs a paid key. Newer models like Command A Reasoning need contact with sales for production use.
Z.ai lists GLM-4.7-Flash, GLM-4.5-Flash and the vision model GLM-4.6V-Flash as free on input, cached input and output.
Anyone with a Z.ai API account
Catch: Only the Flash models above are free; rate limits are not shown on the pricing page. Pricing pages change, so check before building on it. Newer GLM-5.x models are paid.
Kiro's free tier includes 50 credits and Claude Sonnet 4.5 plus open-weight models; first paid upgrade earns $20 in credit.
Anyone signing in with a social login or AWS Builder ID
Catch: Free tier is rate-limited and not available for Enterprise or GovCloud. The $20 bonus applies only to your first paid upgrade, requires a card, and is prorated.
Runway's Free plan includes 125 one-time credits to try its generative video and image tools.
Anyone with a Runway account
Catch: Credits are one-time, not monthly, so they run out quickly on video generation. Watermarks and export limits are not described on the pricing page.
New Google Cloud users with a Gmail account can use Google's Agent Platform APIs free for up to 90 days with an API key and no billing info.
New Google Cloud users with an @gmail.com Google Account
Free tier lasts up to 90 days; quota-limited
Catch: Only for new Google Cloud users with a @gmail.com account. Separate from the $300 Google Cloud Free Program. Existing Google Cloud users, or anyone who enables billing, lose the 90-day free tier and move to paid. Limited to the APIs and quotas express mode supports.
Activating Model Studio (Singapore region) gives a free token quota per model, typically 1M tokens each, valid for 90 days.
New Alibaba Cloud Model Studio users in the Singapore region, international deployment
90-day validity; per-model quota
Catch: You must complete account information to activate. Quota is per model, does not transfer, and expires after 90 days even if unused. It does not cover batch calls, fine-tuning, deployment or storage. After it runs out you are billed pay-as-you-go unless you enable Free Quota Only (off by default).
OpenRouter lists open and partner models at $0 per token; request caps depend on how many credits you have bought.
Anyone with an OpenRouter account
20 req/min - 50 req/day (1,000/day after $10 purchased)
Catch: Free models change often and have tight caps. The daily limit goes up only if you buy credits ($10 or more). Free-model providers can log prompts, so check each model's data policy before sending sensitive data.
build.nvidia.com offers free API access to hosted models for prototyping once you join the NVIDIA Developer Program.
Anyone with a verified NVIDIA account
Catch: Prototyping use only: free-tier calls are rate-limited per model and by current traffic. NVIDIA does not publish one fixed limit and moderators say limits cannot be raised on the free tier (users report roughly 40 requests/min). Production needs paid or self-hosted NIM. Account verification (possibly including a phone number) may be required.
SambaNova's free tier applies when no payment method is linked: 20 requests/min, 20 requests/day and 200K tokens/day per model.
Anyone with a SambaNova Cloud account and no linked payment method
20 RPM - 20 req/day - 200K tokens/day
Catch: Very small daily request cap. The plans page also describes the Free plan as 'add a payment method and purchase credits', so the rules may differ from the rate-limit docs; check your account's Limits page. Not for production.
Kaggle gives free NVIDIA P100 GPU time in notebooks, with a weekly quota of about 30 hours.
Anyone with a Kaggle account
~30 hours/week, resets weekly
Catch: The quota can change with demand. Sessions are time-limited, and the P100 does not speed up non-deep-learning workloads like pandas or scikit-learn.
Colab is free to use and includes access to GPUs and TPUs, though without a guaranteed allowance.
Anyone with a Google account
Catch: Free usage limits and hardware availability change dynamically, with no fixed allowance or guaranteed GPU model. Runtimes get disconnected; web-UI-style workloads (e.g. image-generation front ends) are restricted on the free tier.