On this page
Most "best free AI API" lists copy each vendor's marketing page once and never look again. Free tiers in 2026 do not sit still: in the last few months DeepSeek and Together stopped giving free credit, SambaNova refused our free account without a card, GitHub Models shut down, Cerebras put its trial behind a card, and model IDs that the providers still list started returning 404. This page is built from a data file that we re-check by calling every provider's API, so each row carries the date it was last confirmed.
How to choose
Pick by what your app needs, not by brand. Each line below drops off automatically if that provider's free tier closes or starts asking for a card.
- Fastest replies (chat, voice, agents) → Groq. The fastest generation we have measured on any free tier. (checked 2026-10-11)
- Long documents, PDFs, images → Google Gemini. Flash models are free with a very large context window; expect slow first tokens on thinking models. (checked 2026-10-09)
- Trying many models with one key → OpenRouter. One OpenAI-compatible key for every :free model; the free list changes often, so check the model table. (checked 2026-10-09)
- Code completion and small multilingual models → Mistral. Codestral and the Ministral models answer on a free key; the small and medium models do not. (checked 2026-10-09)
- Large open models (Nemotron, gpt-oss) for free → NVIDIA NIM. Hosts big models free for prototyping, but most of its listed models no longer answer. (checked 2026-10-11)
- Chinese-language apps → Zhipu GLM. Free Flash models with strong Chinese; sign-up needs a phone number. (checked 2026-10-09)
- DeepSeek V4 Flash without paying DeepSeek → Gonka Broker. Free monthly tokens on DeepSeek, GLM and MiniMax models, after a phone check. (checked 2026-10-09)
- AI inside a Cloudflare Worker → Cloudflare Workers AI. Free daily Neurons on the edge; useful if your app already runs on Workers. (checked 2026-10-11)
- Embeddings and reranking for RAG → Cohere. Trial key includes Embed and Rerank; not for production traffic. (checked 2026-10-09)
Free LLM APIs compared
"Free allowance" and "Rate limit" are the provider's own published figures; where a provider does not publish one, we say so rather than repeat a number from another list. "Models that answered" counts the chat models that returned a reply when we called them, not the length of the provider's model list. Use the buttons to narrow the table.
| Provider | Status | Free allowance | Rate limit | Card? | Phone check? | Models that answered | Verified |
|---|---|---|---|---|---|---|---|
| Groq | ✅ Free, works | 1,000 requests/day and 200K tokens/day per model | 30 requests/min, 8K tokens/min per model | No | Not stated | 4 · fastest first token: qwen/qwen3.8-27b 0.88 s |
review due |
| Google Gemini | ✅ Free, works | Not published — shown only in AI Studio | Not published — shown only in AI Studio | No Free tier needs no billing account |
Not stated | 10 · fastest first token: gemini-3.5-flash-lite 1.25 s |
review due |
| Mistral | ✅ Free, works | Not published — shown in your console | Not published; small/medium/magistral models get 0 requests/min on free keys | No | Not stated in the docs | 12 · fastest first token: ministral-8b-latest 0.69 s |
review due |
| OpenRouter | ✅ Free, works | 50 requests/day on :free models; 1,000/day after $10 of credits ever bought | 20 requests/min on :free models | No | Not stated | 8 · fastest first token: nvidia/nemotron-3-super-120b-a12b:free 1.22 s |
review due |
| NVIDIA NIM | ✅ Free, works | Free for Developer Program members, for prototyping; no number published | Not published in the docs | No | Not stated | 13 · fastest first token: nvidia/nemotron-3-super-120b-a12b 0.54 s |
review due |
| Zhipu GLM | ✅ Free, phone check | Free Flash models: GLM-4.7-Flash, GLM-4.5-Flash, GLM-4-Flash-250414 | Concurrency limit per account, shown only in the console | No | Yes Sign-up is by SMS code; overseas numbers accepted |
3 · fastest first token: glm-4-flash-250414 0.65 s |
review due |
| Gonka Broker | ✅ Free, phone check | At least 1M tokens/month, renewed on the 1st | 6 requests/min per account | No The free-tier page says no card; the homepage says 'add a card' |
Yes | 3 · fastest first token: deepseek-ai/DeepSeek-V4-Flash-0731 2 s |
review due |
| Cloudflare Workers AI | ✅ Free, works | 10,000 Neurons/day | Per-model limits; not compared here | Workers Free is the default plan; card rule not stated, and we did not sign up a fresh account to check | Not stated | 6 | review due |
| Cohere | ⏳ Trial only | 1,000 API calls/month on a trial key | 20 chat requests/min (trial key) | Trial key on sign-up; card rule not stated | Not stated | We hold no key: not called by us | review due |
| Cerebras | 💳 Card required | $5 trial credit, expires after 30 days; 1M tokens/day | 5 requests/min (trial) | Yes API stays inactive until a payment method is added |
Not stated | We hold no key: not called by us. Trial tier lists gpt-oss-120b and qwen-3.8-27b | review due |
| SambaNova Cloud | 💳 Card required | Docs list 20 requests/day per model with no payment method; our key got 402 PAYMENT_METHOD_REQUIRED | Docs: 20 requests/min per model (free tier) | Yes | Not stated | Our key now returns 401; not re-tested since 2026-08-31 | review due |
| DeepSeek API | ❌ Closed since 2026-08-31 |
No free credit for new accounts on the pricing page | — | Yes Prepaid balance required |
Not stated | Both models return 402 Insufficient Balance | review due |
| Together AI | ❌ Closed since 2026-08-31 |
No free trial; $15 minimum first purchase, $5 minimum later | — | Yes | Not stated | Account read-only until you deposit | |
| GitHub Models | ❌ Closed since 2026-07-30 |
Service retired | — | Not stated | Not stated | Retired: playground, catalog and inference API shut down | review due |
Model IDs that answered
A model appearing in a provider's /models list is not proof you can call it. On our last pass, NVIDIA listed dozens of models that returned 404 "Function not found", Gemini still listed 2.5-series models that answer "no longer available to new users", and Mistral lists models that free keys are allowed 0 requests per minute on. These are the IDs that actually replied. "First token" is the time until the first visible text of a one-line reply, measured once from our test machine in Asia, some providers through a proxy, so compare rows within a provider rather than across providers. Reasoning models spend most of that time thinking.
| Provider | Model ID that answered | First token | Checked |
|---|---|---|---|
| Groq | qwen/qwen3.8-27b free tier also caps output at 1,000 tokens/min (429 OTPM, 2026-10-11) | 0.88 s | 2026-10-11 |
| Groq | openai/gpt-oss-120b | 1.95 s | 2026-10-11 |
| Groq | allam-2-7b | 4.76 s | 2026-10-11 |
| Groq | openai/gpt-oss-20b | 5.94 s | 2026-10-11 |
| Google Gemini | gemini-3.5-flash-lite | 1.25 s | 2026-10-09 |
| Google Gemini | gemini-3-flash-preview | 1.52 s | 2026-10-09 |
| Google Gemini | gemini-3.6-flash | 2.15 s | 2026-10-09 |
| Google Gemini | gemini-3.5-flash | 2.66 s | 2026-10-09 |
| Google Gemini | gemini-3.1-flash-lite-preview | 2.73 s | 2026-10-09 |
| Google Gemini | gemini-3.1-flash-lite | 2.91 s | 2026-10-09 |
| Google Gemini | gemini-flash-lite-latest | 5.94 s | 2026-10-09 |
| Google Gemini | gemini-flash-latest | 17.56 s | 2026-10-09 |
| Google Gemini | gemini-3.7-flash | 36.58 s | 2026-10-09 |
| Google Gemini | gemini-3.8-flash | 40.79 s | 2026-10-09 |
| Mistral | ministral-8b-latest | 0.69 s | 2026-10-09 |
| Mistral | mistral-code-fim-latest | 0.73 s | 2026-10-09 |
| Mistral | mistral-code-latest | 0.95 s | 2026-10-09 |
| Mistral | voxtral-small-latest | 0.96 s | 2026-10-09 |
| Mistral | codestral-2508 | 0.97 s | 2026-10-09 |
| Mistral | voxtral-small-2507 | 1.27 s | 2026-10-09 |
| Mistral | ministral-8b-2512 | 1.28 s | 2026-10-09 |
| Mistral | ministral-3b-2512 | 2.69 s | 2026-10-09 |
| Mistral | ministral-3b-latest | 2.77 s | 2026-10-09 |
| Mistral | ministral-14b-latest | 3.59 s | 2026-10-09 |
| Mistral | ministral-14b-2512 | 4.08 s | 2026-10-09 |
| Mistral | codestral-latest | 4.98 s | 2026-10-09 |
| OpenRouter | nvidia/nemotron-3-super-120b-a12b:free | 1.22 s | 2026-10-09 |
| OpenRouter | nvidia/nemotron-3-ultra-550b-a55b:free | 1.63 s | 2026-10-09 |
| OpenRouter | liquid/lfm-2.5-2.6b:free | 1.76 s | 2026-10-09 |
| OpenRouter | cohere/north-mini-code:free | 2.9 s | 2026-10-09 |
| OpenRouter | apodex/apodex-1.1-mini:free | 2.99 s | 2026-10-09 |
| OpenRouter | nvidia/nemotron-3.5-lightning:free | 11.35 s | 2026-10-09 |
| OpenRouter | dots-studio/dots-3-note-preview:free | 11.47 s | 2026-10-09 |
| OpenRouter | nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free answered, but the timed streaming call returned no text | not timed | 2026-10-09 |
| NVIDIA NIM | nvidia/nemotron-3-super-120b-a12b | 0.54 s | 2026-10-11 |
| NVIDIA NIM | nvidia/nemotron-3.5-lightning-30b-a3b | 0.64 s | 2026-10-11 |
| NVIDIA NIM | z-ai/glm-5.3-flash | 0.68 s | 2026-10-11 |
| NVIDIA NIM | meta/llama-3.2-11b-vision-instruct | 0.68 s | 2026-10-11 |
| NVIDIA NIM | nvidia/riva-translate-4b-instruct-v2 | 0.69 s | 2026-10-11 |
| NVIDIA NIM | openai/gpt-oss-20b | 0.73 s | 2026-10-11 |
| NVIDIA NIM | nvidia/ising-calibration-1.5-31b | 0.76 s | 2026-10-11 |
| NVIDIA NIM | meta/muse-glimmer-30b | 1.3 s | 2026-10-11 |
| NVIDIA NIM | nvidia/nemotron-3-ultra-550b-a55b | 1.31 s | 2026-10-11 |
| NVIDIA NIM | deepseek-ai/deepseek-v4.1-flash | 1.38 s | 2026-10-11 |
| NVIDIA NIM | nvidia/nemotron-3-nano-omni-30b-a3b-reasoning | 1.45 s | 2026-10-11 |
| NVIDIA NIM | google/diffusiongemma-26b-a4b-it | 8.41 s | 2026-10-11 |
| NVIDIA NIM | google/gemma-4-31b-it answered on a retry; the timed streaming call was dropped | not timed | 2026-10-11 |
| Zhipu GLM | glm-4-flash-250414 | 0.65 s | 2026-10-09 |
| Zhipu GLM | glm-4.5-flash | 5.22 s | 2026-10-09 |
| Zhipu GLM | glm-4.7-flash answered once; the timed call got 429 | not timed | 2026-10-09 |
| Gonka Broker | deepseek-ai/DeepSeek-V4-Flash-0731 | 2 s | 2026-10-09 |
| Gonka Broker | MiniMaxAI/MiniMax-M2.7 | 2.61 s | 2026-10-09 |
| Gonka Broker | zai-org/GLM-5.3-Flash | 12.59 s | 2026-10-09 |
| Cloudflare Workers AI | @cf/meta/llama-3.1-8b-instruct answered; 579 tokens for 20.4 neurons (~490 such replies/day), 5.5 s non-streaming | not timed | 2026-10-11 |
| Cloudflare Workers AI | @cf/meta/llama-3.2-3b-instruct answered; 536 tokens for 16.6 neurons, 3.1 s non-streaming | not timed | 2026-10-11 |
| Cloudflare Workers AI | @cf/openai/gpt-oss-20b answered; 600 tokens for 17.9 neurons, 7.2 s non-streaming | not timed | 2026-10-11 |
| Cloudflare Workers AI | @cf/meta/llama-3.3-70b-instruct-fp8-fast answered on the free plan; 578 tokens for 119.7 neurons (~83/day) | not timed | 2026-10-11 |
| Cloudflare Workers AI | @cf/google/gemma-4-26b-a4b-it answered; 600 tokens for 16.7 neurons | not timed | 2026-10-11 |
| Cloudflare Workers AI | @cf/zai-org/glm-4.7-flash answered; 600 tokens for 22.0 neurons | not timed | 2026-10-11 |
Speed: tokens per second, measured 2026-08-31
The table above answers "can I call it". This one answers "how fast is it once it starts". We sent the same ~250-word prompt to each free API with one streaming request (temperature 0.3, 500-token cap) from a US GitHub Actions runner, and computed generation speed from each API's own usage token count, excluding the time to first token. These numbers are from 2026-08-31 and have not been re-run since; the model IDs marked as changed no longer answer the same way today.
| Provider | Model tested | First token | Generation speed | Since then |
|---|---|---|---|---|
| Groq | openai/gpt-oss-120b | 0.38 s | ~521–527 tok/s | Still answers |
| Mistral | mistral-small-latest | 0.43 s | ~161–179 tok/s | Free keys now get 0 requests/min on this model |
| Google Gemini | gemini-3.6-flash | 17–34.5 s (thinking) | 141 tok/s | See the model table for what answers now |
| OpenRouter | nvidia/nemotron-3-super-120b-a12b:free | 2.6–2.7 s | ~38–77 tok/s | Still answers |
| NVIDIA NIM | nvidia/nemotron-3-super-120b-a12b | 3.5–7.6 s | ~26–100 tok/s, wildly inconsistent between runs | Still answers |
| Zhipu GLM | glm-4-flash | 2.6 s | ~20–23 tok/s | Still answers, but is no longer in GLM's model list; the listed free ID is glm-4-flash-250414 |
Groq was not slightly faster; it was about 3× the next one. If generation speed matters at all, that settles it.
Re-tested 2026-10-11
We re-ran the two fastest free APIs and the Nemotron comparison on 11 October 2026, same prompt and method, but from a machine in Asia through a proxy rather than the US runner — so read the absolute numbers as rough and the direction as the finding.
| Provider and model | Runs | Generation speed | Change since 2026-08-31 |
|---|---|---|---|
Groq openai/gpt-oss-20b | 3 | 813–885 tok/s (median 839) | Still the fastest free model we have measured |
Groq openai/gpt-oss-120b | 3 | 261–454 tok/s (median 353) | Slower than August's ~521–527 and less steady |
Groq qwen/qwen3.8-27b | 3 | 445–563 tok/s (median 544) | New; free tier also caps output at 1,000 tokens/min |
NVIDIA NIM nvidia/nemotron-3-super-120b-a12b | 2 | 136.5 and 200.5 tok/s | Up from ~26–100 tok/s |
OpenRouter nvidia/nemotron-3-super-120b-a12b:free | 1 (2nd dropped by our proxy) | 38.2 tok/s, only ~36 visible tokens | Down from 45.7 tok/s |
Groq still leads, but the gap to NVIDIA's own endpoint has narrowed from 3× to roughly 2–3× on these runs.
In August, the same model was faster through OpenRouter than from NVIDIA itself — in October it was the reverse (200.5 vs 38.2 tok/s, table above). NVIDIA's nemotron-3-super-120b is served both by NVIDIA's own free API and by OpenRouter's free tier. Same model, same prompt, same runner on 2026-08-31:
| Serving nemotron-3-super-120b | Time to first token | Generation speed |
|---|---|---|
| OpenRouter (free) | 2.7 s | 45.7 tok/s |
| NVIDIA (own API) | 7.6 s | 36.7 tok/s |
Free capacity is shared and unevenly provisioned, and it moves: neither the first-party endpoint nor a router is reliably the faster route to a model, so time both before you commit.
How we check
- Can it be called: for every provider we hold a free key for, we fetch the
/modelslist and call each chat model once with a one-line prompt. A row only counts as answered on an HTTP 200 with a reply. Connection drops on our side are retried and are never counted as a model failure. - Rate limits that look like errors: a Mistral 429 is checked against the
x-ratelimit-limit-req-minuteheader. A limit of 0 means the free tier does not include that model, which is different from "try again later". - First token: one streaming request per model that answered, timing the first visible text.
- Allowances, limits, card and phone rules: from each provider's own pricing and rate-limit pages, linked from the data file. Gemini and Mistral do not publish free-tier limits (they show them only inside your console), so we do not print one.
- Providers we have no working key for (Cerebras, Cohere, Cloudflare Workers AI, and SambaNova since our old key started returning 401) are described from their official pages and our earlier tests only, and the table says so.
- The benchmark harness is open source: free-ai-api-benchmark.
One code pattern for almost all of them
Groq, Gemini, Mistral, OpenRouter, NVIDIA NIM, GLM and Gonka Broker all accept the OpenAI chat format, so switching provider means changing the base URL, the key and the model ID. Take the model ID from the "Model IDs that answered" table above, not from an old tutorial.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_KEY",
base_url="https://openrouter.ai/api/v1", # swap per provider
)
response = client.chat.completions.create(
model="nvidia/nemotron-3-super-120b-a12b:free", # an ID from the table above
messages=[{"role": "user", "content": "What is the capital of France?"}],
)
print(response.choices[0].message.content)
Base URLs: Groq https://api.groq.com/openai/v1, Gemini https://generativelanguage.googleapis.com/v1beta/openai, Mistral https://api.mistral.ai/v1, NVIDIA https://integrate.api.nvidia.com/v1, GLM https://open.bigmodel.cn/api/paas/v4, Gonka Broker https://proxy.gonkabroker.com/v1. Cloudflare Workers AI runs inside a Worker via env.AI.run(...) or over its REST API.
Frequently asked questions
Can I use several free APIs in one app?
Yes, and with free tiers this unstable you should keep at least one fallback. Because the call shape is shared, a router such as LiteLLM can fail over from one provider to the next when a free quota runs out.
Why do some rows say "Not published"?
Because the provider does not publish that number. Google and Mistral show free-tier limits only inside your own console, and they change them without notice. Other lists quote a figure anyway; we would rather tell you where to look.
Is DeepSeek still free?
The DeepSeek models are, on other providers: check the model table for DeepSeek IDs on NVIDIA NIM and Gonka Broker. DeepSeek's own API is not: it returned 402 Insufficient Balance to a free account, see the table for the date.
Related reads
- Groq vs Cerebras vs Gemini Free API: Limits and Speed, Tested — head-to-head speed tests
- OpenRouter free models — when model variety matters more than one provider
- Groq Free Tier Gives You Throughput, Not Context — what breaks when a free key meets a real agent
- Cohere Free API: Embedding and Rerank for RAG — pair with any chat API for a free RAG stack