Most "best free AI API" lists just re-copy each vendor's marketing page. We did something different: we wrote a small benchmark harness, pointed it at every free API we could get a key for, and measured them ourselves — real generation speed from each API's own token counts, on the same US runner, same prompt, same day. The results are not what the listicles say.
The collapse: half of them now want your money
The single biggest finding isn't a speed number — it's how many "free" tiers have quietly closed. Every one of these turned us away at the door in late August 2026:
| Platform | What we hit |
|---|---|
| DeepSeek | 402 Insufficient Balance — the API now needs a funded balance |
| SambaNova | 402 PAYMENT_METHOD_REQUIRED, balance 0 — a card is now mandatory |
| Together AI | Account is in read-only mode until you "make an initial deposit" |
| Cerebras | Free trial now requires a verified payment method up front |
| xAI (Grok) | Free credits require a card on file |
| GitHub Models | 410 github_models_retirement_brownout — gone entirely, retired July 30, 2026 |
GitHub Models deserves its own line: it didn't get more expensive, it ceased to exist. Calls now return 410 github_models_retirement_brownout — GitHub fully retired the service on July 30, 2026. That's a shame: it was the lowest-friction free API there was, since any GitHub account already had access. If you built on it, you've already been migrated off whether you noticed or not — see our GitHub Models retirement guide for what to move to.
On top of that, the models moved under everyone's feet: Groq quietly dropped Llama 3.3 (it's Enterprise-only now; the free catalog is gpt-oss and Qwen), NVIDIA's Llama-Nemotron 70B returns "not found for account", and Google retired Gemini 2.0/2.5 for a 3.x line. If you copy a 2025 tutorial, the model string in line one no longer exists. This is why we re-verify everything instead of quoting docs.
What's actually still free — and how fast it really is
Here's the genuinely-free-with-no-card set, measured on the same US runner on 2026-08-31. "Generation speed" is throughput after the first token, computed from each API's real usage token count — the number that reflects the model's raw speed rather than your distance to the server.
| Platform | Model tested | Generation speed | Note |
|---|---|---|---|
| Groq | gpt-oss-120b | ~521–524 tok/s | Fastest by a wide margin, and the most consistent; LPU hardware |
| Mistral | mistral-small | ~161–179 tok/s | Solid all-rounder, low latency |
| OpenRouter | nemotron-3-super-120b (free) | ~46–77 tok/s | Free models can hit upstream rate limits |
| NVIDIA | nemotron-3-super-120b | ~26–100 tok/s | Wildly inconsistent between runs |
| GLM (Zhipu) | glm-4-flash | ~20–23 tok/s | Genuinely free, but the slowest here |
| Gemini | gemini-3.6-flash | see below | A reasoning model — a special case |
The Gemini asterisk: gemini-3.6-flash is a thinking model. Across our runs it spent 17–30 seconds "thinking" before the first visible token, then emitted a short answer. Raw tokens-per-second is the wrong lens for it — you're paying latency for reasoning quality, not for throughput. Great for hard problems, wrong choice for a snappy chatbot.
The surprise: OpenRouter beats NVIDIA at NVIDIA's own model
Because free platforms rarely host the same model any more, a fair head-to-head is hard. But NVIDIA's nemotron-3-super-120b is on both NVIDIA's own API and OpenRouter's free tier — the identical model. Same model, same prompt, same runner:
| 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 |
OpenRouter served NVIDIA's model both faster and with a third of the latency of NVIDIA's own free endpoint. Counter-intuitive, but that's what routing + spare capacity buys you. (NVIDIA's own numbers also swung a lot between runs, from ~37 to ~100 tok/s — its free tier is the least predictable of the bunch.)
How we measured it (so you can trust the numbers)
- One streaming request per API, temperature 0.3, 500-token cap, the same ~250-word prompt.
- Generation speed =
completion_tokens ÷ generation time, excluding time-to-first-token, using each API's ownusagecount — not a character estimate. - Run from a US GitHub Actions runner so location doesn't skew latency; measured 2026-08-31.
- Your numbers will vary with prompt, model, load, and network. We re-run rather than quote vendor claims.
Who should use what
- Need raw speed for a real-time chat or agent? Groq, no contest (~520 tok/s, no card).
- Want a balanced free workhorse? Mistral is fast and low-latency.
- Want a big model for free? nemotron-3-super-120b — and get it via OpenRouter, not NVIDIA's own API.
- Hard reasoning, willing to wait? Gemini's thinking model.
- Don't bother (unless you'll pay): DeepSeek, SambaNova, Together, Cerebras, xAI — the free door is closed.
All figures first-hand from our open benchmark harness, run 2026-08-31. Free tiers change weekly in 2026 — we'll re-run and update rather than let this go stale.