· CostMyTokens · Analysis · 8 min read
Free LLM API Tiers Compared: What's Actually Free in 2026
One genuine free tier among the major model labs, several one-time credits, gateway free lines for open models, and two famous offers that no longer exist — every major LLM API's free access, checked on 2026-10-26.
“Free” is the most abused word in API pricing. Of the nine providers below, exactly one major model lab runs a genuine, ongoing free tier for its own models (Google Gemini). Two gateways (Groq, OpenRouter) also offer ongoing free inference — but only for open-weight / free-roster models, not frontier closed models. The rest offer a one-time signup credit, bundle credits into a paid subscription, or quietly ended their free program while old guides keep recommending it. Every entry here was checked on 2026-10-26 (Asia/Shanghai) against the provider’s own documentation where it exists, and against current trackers where it does not — and each row says which. Free tiers change without notice; treat the check date as part of the data.
Disclosure: this article contains no affiliate links. Where CostMyTokens does use affiliate links elsewhere, they never affect order — every table is sorted by the terms offered, not by commission. Read our disclosure.
The comparison
| Provider | What you get | Type | Limits / amount | Card required | Status (checked 2026-10-26) |
|---|---|---|---|---|---|
| Google Gemini API (AI Studio) | Free calls to eligible Flash / Flash-Lite models | Ongoing free tier | Google no longer publishes one central free-quota table; live RPM / TPM / RPD are in AI Studio. Trackers of that console (2026-10-25) report Flash-Lite ~10 RPM / 250K TPM / 500 RPD and Flash ~5 RPM / 250K TPM / 20 RPD | No | Active — the only ongoing free tier among major model labs |
| Anthropic (Claude API) | No published free API tier | None (prepaid) | Official help: buy prepaid credits before use. Max / Team subscriptions can claim monthly API credits after linking Console. Older third-party reports of a ~$5 test credit are unverified on current signup | Yes (typical Console signup) | No free tier — check your Console for any account-specific grant |
| OpenAI API | Nothing ongoing | None | The “Free” usage tier in OpenAI’s docs is a $100/month usage ceiling, not funds | — | No free tier and no published signup credit |
| xAI (Grok API) | Nothing ongoing | None (ended) | xAI billing docs (updated 2026-10-10) list prepaid credits and promo codes only; the $150/month data-sharing credit no longer appears | — | Prepaid pay-as-you-go only |
| DeepSeek API | Possible one-time granted balance | One-time credit (unpublished) | Docs describe a granted balance spent before topped-up balance, with no published amount, eligibility or expiry | No | No free tier; every call beyond any granted balance is billed |
| Kimi API (Moonshot) | ¥15 voucher on the China platform | One-time credit | China platform only, requires a mainland-China mobile number, valid 3 months. International (platform.kimi.ai): no free signup credit; $1 minimum top-up to start | No | Regional — most readers will not qualify |
| Novita AI (gateway) | New-user voucher (amount not published in current docs) | One-time credit | Current docs confirm a new-user voucher without naming the amount. An older Novita guide (2024) said $0.50. Official referral is Give $10 / Earn $10 LLM credits. Third-party “$10” claims usually describe referral or promos, not a standing free grant | No | No ongoing free tier |
| Groq (gateway) | Free inference of open models | Ongoing, rate-limited | Groq console rate-limit docs list open models such as GPT OSS / Qwen at 30 RPM, 1,000 RPD, 8K TPM (verify live in your org Limits page) | No | Active — open models only |
| OpenRouter (gateway) | Free-model roster (:free) | Ongoing, rate-limited | Official: 20 RPM; 50 requests/day if lifetime purchases under $10; 1,000/day after ≥ $10 purchased | No | Active — free models only; frontier models are pass-through paid |
The only lab free tier: Google Gemini
Google’s AI Studio free tier is what most people imagine a free tier to be: no card, no expiry date, real models, and requests that simply stop (HTTP 429) rather than bill you when you hit the cap. Two things define it in practice. First, the allowance is per model and project, and Google no longer publishes one central free-quota table — your project’s live quotas are in the AI Studio rate-limit console (same caveat as our Gemini pricing page). Trackers of that console still show a wide gap: Flash-Lite models around 500 requests/day versus newer Flash models around 20 (source). Second, the price of free is your data: on the free tier, Google may use your inputs to improve its products (paid tier: no), which rules it out for sensitive or production user data. And if you enable billing on a project, that project leaves the free tier entirely — keep prototyping on a separate project. Official pricing confirms free-of-charge input/output for eligible Flash and Flash-Lite models (pricing).
One-time credits: Anthropic, Kimi, Novita
Anthropic is the most misreported. Its help center is clear that Claude API usage is prepaid — buy credits before you call the API (help). There is no published free API tier for new Console accounts. Separately, Claude Max and Team subscriptions can claim monthly API credits after linking a Console organization — that is a paid-plan feature, not a free tier. Treat any small test credit you see in Console as account-specific and unverified for new signups in 2026; do not plan around a “$5 free Claude API” blog post.
Kimi publishes its voucher properly: ¥15 for new users who register with a mainland-China mobile number, valid three months, applied automatically (help). The international platform requires at least a $1 top-up before API use (limits).
Novita AI, the gateway on our provider pages, documents a new-user voucher but does not publish the dollar amount in current docs. An older Novita blog (2024) mentioned $0.50 on registration; treat that as historical unless your billing page shows it. The standing official free-ish path for LLM credits is the referral program (Give $10, Earn $10). Larger “$10” headlines on third-party sites usually describe that referral or a short promo — not an unlimited free tier.
The two famous offers that are gone
OpenAI’s “$5 free credits.” Still quoted by most comparison pages, but OpenAI’s rate-limits guide and prepaid billing help (read 2026-10-26) find no signup credit: the “Free” usage tier is a $100/month spending ceiling for unpaid organizations, and a ceiling cannot pay for a request (rate limits; prepaid billing). New API accounts use prepaid billing with a $5 minimum purchase. Requests are free only if your account happens to hold a credit grant from a program or promotion. If your billing page shows a grant, spend it before it expires; if it does not, no guide can conjure one.
xAI’s “$150 a month.” The data-sharing credit program was real — and it is the single most recycled dead offer in this space. xAI’s current billing documentation (updated 2026-10-10) lists prepaid credits and promo codes only; the $150/month program no longer appears (billing). The only free xAI inference left is the consumer playground, not the API.
No free tier, stated plainly: DeepSeek
DeepSeek does not pretend otherwise. Its docs describe a “granted balance” that is spent before your topped-up balance but publish no amount, eligibility rules or expiry — so any signup credit is a bonus, not a plan (pricing; balance API exposes granted_balance). The chat app is free; the API is pay-as-you-go from there, at prices low enough that the distinction matters less than elsewhere — see the DeepSeek pricing breakdown and the DeepSeek provider page. The models are also open-weight, so “free” via self-hosting is an infrastructure decision, not an API tier.
Gateway free access: Groq and OpenRouter
If the goal is free calls rather than a specific frontier model, gateways serving open-weight models are the practical route. Groq publishes organization rate limits in console docs — open models such as GPT OSS and Qwen commonly sit at 30 RPM / 1,000 RPD / 8K TPM (rate limits); always confirm the Limits page for your org, because the binding constraint on long prompts is usually tokens-per-minute. OpenRouter keeps a roster of :free models behind one key: official free-usage limits are 20 requests/minute and 50/day until you have purchased at least $10 lifetime, then 1,000/day (limits). Both are real ongoing free access — for open / free-roster models only, not a substitute for Gemini’s free Flash tier or a paid frontier API.
What a free tier actually buys you
A free tier is a rate limit, not a budget. 500 requests a day is a generous prototype allowance and a hard ceiling for anything public: one busy hour can spend the day. Tokens-per-minute binds before requests-per-day on long-context work — at 250K TPM, a handful of 100K-token prompts per minute is the real cap. And every free arrangement has a non-price price: your data trains someone’s model (Google’s free tier, xAI’s old program), your throughput is last in line, or your credit expires on a clock you did not set. None of that makes free tiers useless — they are the right way to evaluate an API and to run personal projects. It makes them a starting point with a known exit.
When free ends
The exit is cheaper than it used to be. At production volumes the paid ladder now starts at $1.80 per 10M tokens (GPT-6 Luna, standard mix) in our OpenAI pricing overview, and the cheapest LLM API ranking shows the full field at the same workload. When a free cap starts costing you users, that is the table to price the move with — and the calculator will cost your exact mix before you commit.
Figures checked 2026-10-26 (Asia/Shanghai). Free tiers and signup credits change without notice — verify in each provider’s console before you build on them.