Guide · verified 2026-10-15 · no signup
Free LLM APIs in 2026: which are actually free
OpenRouter, Groq and SambaNova are the only providers with a genuinely free, no-expiry API tier that reaches GPT-4o-class (100B+) models — and all three work with no credit card. Everything else is either caveated (tight caps, eval-only licences, or training on your data) or trial-only. Here is the honest breakdown, verified 2026-10-15.
The short version
- Actually free, no card, no expiry: OpenRouter, Groq, SambaNova. All three serve 100B+ open models (gpt-oss-120B, Llama-3.3-70B, DeepSeek V3.1, Kimi K2) for free within rate limits.
- Free but with a real catch: Mistral (eval-only, org-wide monthly cap), Cohere (1,000 calls/month total), Google AI Studio (trains on your data; text/audio only).
- Not a free tier — trial credits: Cerebras ($5/30-day, card required), Scaleway (one-time grant), NVIDIA build (expiring credits), Together / Fireworks / Nebius (~$1), Hugging Face ($2/mo behind PRO). GitHub Models was retired 30 Jul 2026.
Every free LLM API, compared
Tier means: Standing = free forever within limits · Constrained = free but seriously capped or caveated · Trial = one-time or expiring credits.
| Provider | Models | Free? | RPM | Daily cap | Token cap | Card? |
|---|---|---|---|---|---|---|
| OpenRouter | Inkling 975B · Nemotron 3 Ultra 550B · Dots3-Note 280B · Inkling Small 276B + 10 more | Standing | 20 | 50 req/day (free) → 1,000 req/day after a one-time ≥$10 credit (lifetime) | No token/day cap | No |
| Groq | gpt-oss-120B · Llama-3.3-70B · Kimi K2 ~1T · Qwen3-32B · Llama-4 + more | Standing | ~30 (free) → ~1,000 with a card | ~1,000 req/day (free); daily cap removed on the card-backed Developer tier | Low TPM: ~6K–30K tokens/min per model | No |
| SambaNova Cloud | Llama-3.3-70B · DeepSeek V3.1 · DeepSeek V3.2 · gpt-oss-120B · Gemma 4 31B | Standing | 20 | 20 req/day per model | 200K tokens/day per model | No |
| Mistral La Plateforme | Mistral Large ~123B · Medium 3.x · Small 3.x · Codestral · Devstral | Constrained | ~60 (≈1 req/s) | No daily cap | ~1B tokens/month — shared ACROSS YOUR WHOLE ORG; ~500K TPM | No |
| Cohere | Command A ~111B · Command R+ ~104B · Command R ~32B · Rerank · Embed | Constrained | 20 (Chat) | No daily cap | 1,000 API calls/MONTH — total, across the whole account | No |
| Google AI Studio / Gemini | Gemini 2.5 Pro / Flash / Flash-Lite · 3.x Flash · Gemma 4 · TTS / Live / embeddings | Constrained | ~5–15 (unofficial, console-only) | ~250–1,500 req/day (unofficial) | ~250K TPM (unofficial) | No |
| Cerebras | Qwen3-235B / 480B · Llama-3.3-70B · gpt-oss-120B | Trial | 5 | — | 1M tokens/day (during the trial) | Yes |
| Scaleway | Llama-3.3-70B · Qwen3-235B · DeepSeek-R1-distill-70B · gpt-oss-120B · Mistral-Medium | Trial | 60 | — | One-time 1M-token grant (non-recurring), then paid | Yes |
| NVIDIA build (build.nvidia.com) | Llama-3.1-405B · DeepSeek-R1 671B · Nemotron-4 340B · Qwen2.5-72B | Trial | ~40 | — | Expiring trial credits | No |
Also trial-only, not tabled: Together AI · Fireworks AI · Nebius · Hugging Face Inference · DeepInfra · GitHub Models.
Actually free
A standing free tier with no expiry. You can keep calling 100B+ models for free, within rate limits, forever.
An aggregator that exposes a rotating set of 14 genuinely-free (:free) model variants from many labs behind one OpenAI-compatible API. No card needed for the 50 req/day tier; a single lifetime ≥$10 top-up lifts you to 1,000 req/day. Bring-your-own-key (BYOK) routing is ungated.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Inkling | 975B / 41B active | 1M | largest free text model |
| Nemotron 3 Ultra | 550B / 55B active | 1M | — |
| Dots3-Note | 280B / 16B active | 512K | — |
| Inkling Small | 276B / 12B active | 1M | — |
| Nemotron 3 Super | 120B / 12B active | 256K | — |
| Laguna S 2.1 | 118B / 8B active | 256K | code |
| Nemotron 3.5 Lightning | 30B / 3B active | — | — |
| Nemotron Nano Omni | 30B / 3B active | — | — |
| Gemma 4 | 31B | — | — |
| Gemma 4 | 26B / 4B active | — | — |
| North Mini Code | 30B / 3B active | — | code |
| Laguna XS | 33B / 3B active | — | — |
| LFM2.5 | 2.6B | — | — |
| Nemotron Content-Safety | 4B | — | moderation |
⚠ The two Lyria "free" entries are music models that actually bill $0.08/song — not free. They are excluded from the 14 above.
The fastest free tier for open models (custom LPU hardware). No card for the free tier; adding one unlocks the Developer tier (~1,000 RPM, no daily cap). The free TPM is tight, so it suits short, fast calls more than long documents.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| gpt-oss-120B | 120B / 5.1B active | — | OpenAI open-weight |
| Llama-3.3-70B | 70B | — | — |
| Qwen3-32B | 32B | — | — |
| Llama-4 Scout | 109B / 17B active | — | — |
| Llama-4 Maverick | ~400B / 17B active | — | throttled: 15 RPM / 500 req/day |
| Kimi K2 | ~1T / 32B active | — | largest free model anywhere |
| DeepSeek-R1-distill-70B | 70B | — | reasoning |
| gpt-oss-20B | 20B | — | — |
| Llama-3.1-8B-instant | 8B | — | — |
A standing free tier on custom RDU hardware, with high throughput. The catch is volume: 20 requests and 200K tokens per day, counted per model — generous for evaluating, thin for production. No card required.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Llama-3.3-70B | 70B | — | — |
| DeepSeek V3.1 | 671B / 37B active | — | — |
| DeepSeek V3.2 | 671B / 37B active | — | preview |
| gpt-oss-120B | 120B / 5.1B active | — | — |
| Gemma 4 31B | 31B | — | preview |
Free but limited / caveated
Perpetual free access, but with a real catch — a tight monthly cap, an eval-only licence, or your inputs being used to train the model.
Mistral's free "Experiment" tier needs no card, but you do verify a phone number, and the licence is evaluation-only (not for production). The ~1B tokens/month allowance is shared across every key in your organization, not per key.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Mistral Large | ~123B | — | — |
| Mistral Medium 3.x | undisclosed | — | — |
| Mistral Small 3.x | 24B | — | open weights |
| Codestral | 22B | — | code |
| Devstral | 24B | — | agentic code |
⚠ Experiment tier is eval-only — read Mistral's terms before shipping anything on it.
Cohere's trial keys never expire, but the hard ceiling is 1,000 calls per month for the entire account — fine for prototyping a retrieval or rerank pipeline, far too low for anything live.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Command A | ~111B | — | — |
| Command R+ | ~104B | — | — |
| Command R | ~32B | — | — |
| Rerank | — | — | reranking |
| Embed | — | — | embeddings |
⚠ The 1,000 calls/month cap is account-wide, not per key.
A genuinely useful free tier covering Gemini's text, reasoning, audio and embedding models — but with a serious caveat: Google uses free-tier inputs and outputs to train its models (the paid tier does not). Limits are no longer officially published; the figures here are community-observed.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Gemini 2.5 Pro | undisclosed | 1M+ | reasoning |
| Gemini 2.5 Flash / Flash-Lite | undisclosed | 1M | — |
| Gemini 3.8 / 3.6 / 3.5 / 3.1 Flash & Flash-Lite | undisclosed | 1M | — |
| Gemma 4 | open weights | — | — |
| TTS / Live / embeddings | — | — | text, audio & embeddings only |
⚠ Free tier is TEXT / reasoning / audio / embeddings ONLY — image, video and music generation are NOT free. Google TRAINS on free-tier data; commercial use is allowed but your inputs may be used for training.
Trial credits only — not really free
A one-time grant or expiring credits. Useful for a weekend, but not a free tier you can build on. Several require a card.
Blistering speed and a strong catalog, but it is a $5 / 30-day trial and a credit card is required to start. Not a standing free tier — the credit runs out.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Qwen3-235B | 235B / 22B active | — | — |
| Qwen3-480B | 480B / 35B active | — | code |
| Llama-3.3-70B | 70B | — | — |
| gpt-oss-120B | 120B / 5.1B active | — | — |
⚠ Card required · $5 credit · expires after 30 days.
EU-hosted inference with a one-time 1M-token grant. It does not recur, so once the grant is spent you are on pay-as-you-go. Good if EU data residency matters; not a free tier.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Llama-3.3-70B | 70B | — | — |
| Qwen3-235B | 235B / 22B active | — | — |
| DeepSeek-R1-distill-70B | 70B | — | reasoning |
| gpt-oss-120B | 120B / 5.1B active | — | — |
| Mistral-Medium | undisclosed | — | — |
⚠ One-time 1M-token grant, non-recurring · EU-hosted · then paid.
A playground with the biggest open models, but it runs on expiring trial credits — a preview, not a free tier you can depend on.
| Free model | ~Params | Context | Notes |
|---|---|---|---|
| Llama-3.1-405B | 405B | — | — |
| DeepSeek-R1 | 671B / 37B active | — | reasoning |
| Nemotron-4 340B | 340B | — | — |
| Qwen2.5-72B | 72B | — | — |
⚠ Expiring trial credits — not a free tier.
The rest: trial credits and dead ends
These turn up on "free LLM API" lists but are not standing free tiers — treat them as short-lived credit or already gone:
- Together AI: ~$1 of trial credits, then paid.
- Fireworks AI: ~$1 of trial credits, then paid.
- Nebius: ~$1 of trial credits, then paid.
- Hugging Face Inference: $2/month of credit — but only behind the $9/month PRO plan (tiny).
- DeepInfra: No free tier.
- GitHub Models: RETIRED 30 Jul 2026 — the free playground is dead.
Compare these free models side by side in Verna One
Register your OpenRouter, Groq or SambaNova key once, then run one prompt across all of them at once — compare quality, latency and output in a single view, with automatic fallback when a free tier rate-limits you. The workbench for AI.
Try Verna One free →Methodology & caveats
- Last verified 2026-10-15 against each provider's live documentation and dashboard. Free tiers change often — re-check before you commit.
- GPT-4o size is an estimate. OpenAI has never disclosed GPT-4o's parameter count; the ~200B figure is a community mixture-of-experts estimate. The open-model parameter counts here come from provider and Hugging Face model cards. "Active" params are the experts fired per token on MoE models.
- Google trains on free-tier data. Inputs and outputs on the free Gemini / AI Studio tier may be used to improve Google's models; the paid tier does not. Do not send anything sensitive through the free tier.
- Unofficial limits are marked. Where a provider moved its limits into the console (Google), the numbers are community-observed and should be treated as approximate.
- We sell a router, and this page still lists the free tiers that compete with paying us — labelled for what they are.
Frequently asked
What's the best free LLM API with no credit card?
OpenRouter, Groq and SambaNova Cloud are the only providers offering a standing, no-expiry free tier that reaches GPT-4o-class (100B+) models without a credit card. OpenRouter gives 50 requests/day (lifting to 1,000/day after a one-time ≥$10 credit), Groq offers the fastest inference at roughly 30 requests/minute and ~1,000/day, and SambaNova allows 20 requests and 200K tokens per model per day.
Is there a free GPT-4o alternative API?
Yes. GPT-4o-class quality is available free via gpt-oss-120B, Llama-3.3-70B, DeepSeek V3.1 and Qwen3-32B on Groq, SambaNova and OpenRouter. OpenAI has never disclosed GPT-4o's parameter count (community estimates put it around 200B, mixture-of-experts), so these are capability comparisons, not size-for-size ones.
What are OpenRouter's free rate limits?
OpenRouter free (:free) models allow 20 requests/minute and 50 requests/day with no credit card. A single one-time credit of ≥$10 raises the daily limit to 1,000 requests/day for the lifetime of the account. Bring-your-own-key routing is not rate-gated.
Which free LLM providers require a credit card?
Among genuinely free tiers, none do — OpenRouter, Groq, SambaNova, Mistral, Cohere and Google AI Studio all start without a card. A card is only needed for trial-credit offers such as Cerebras ($5 for 30 days) and some paid upgrade tiers.
Does Google's free Gemini tier train on my data?
Yes. Google uses inputs and outputs from the free tier of AI Studio and the Gemini API to improve its models; the paid tier does not. Commercial use of the free tier is allowed, but if your prompts contain anything sensitive, use the paid tier or another provider.
Is Google AI Studio free for image and video generation?
No. The free Gemini tier covers text, reasoning, audio and embeddings only. Image, video and music generation are not included in the free tier and are billed.
Are trial credits (Cerebras, NVIDIA, Together) the same as a free tier?
No. Trial credits are one-time or expiring — Cerebras is $5 for 30 days (card required), Scaleway is a one-time 1M-token grant, NVIDIA build runs on expiring credits, and Together, Fireworks and Nebius give about $1 each. They are good for a quick evaluation but you cannot build on them long-term.
What is the single most capable model I can call for free?
Through Groq you can call Kimi K2 (~1 trillion parameters, 32B active) for free within the rate limits, and through OpenRouter the Inkling 975B model. Both are large mixture-of-experts models in the GPT-4o capability class.