Free LLM APIs in July 2026: what's actually available
OpenKey team ·
As of today, OpenKey routes 25 models at zero cost per token. That’s the whole list — not a marketing rounding of “20+” or “dozens,” but a countable set you can check against /models/free right now. Here’s what’s actually on it, and the caveats you need before you build anything that depends on a free model staying free.
The list
| Lab | Model | Context window |
|---|---|---|
| Meta | Llama 3.3 70B Instruct (free) | 131,072 |
| Meta | Llama 3.2 3B Instruct (free) | 131,072 |
| Qwen | Qwen3 Coder 480B A35B (free) | 1,048,576 |
| Qwen | Qwen3 Next 80B A3B Instruct (free) | 262,144 |
| NousResearch | Hermes 3 405B Instruct (free) | 131,072 |
| OpenAI | gpt-oss-120b (free) | 131,072 |
| OpenAI | gpt-oss-20b (free) | 131,072 |
| Gemma 4 31B (free) | 262,144 | |
| Gemma 4 26B A4B (free) | 262,144 | |
| NVIDIA | Nemotron 3 Ultra (free) | 1,000,000 |
| NVIDIA | Nemotron 3 Super (free) | 1,000,000 |
| NVIDIA | Nemotron 3 Nano 30B A3B (free) | 256,000 |
| NVIDIA | Nemotron 3 Nano Omni (free) | 256,000 |
| NVIDIA | Nemotron Nano 12B 2 VL (free) | 128,000 |
| NVIDIA | Nemotron Nano 9B V2 (free) | 128,000 |
| NVIDIA | Nemotron 3.5 Content Safety (free) | 128,000 |
| Poolside | Laguna M.1 (free) | 262,144 |
| Poolside | Laguna XS.2 (free) | 262,144 |
| Liquid | LFM2.5-1.2B-Instruct (free) | 32,768 |
| Liquid | LFM2.5-1.2B-Thinking (free) | 32,768 |
| Cohere | North Mini Code (free) | 256,000 |
| Cognitive Computations | Uncensored (free) | 32,768 |
That’s 22 text models. Three more entries carry is_free: true but aren’t general-purpose LLMs worth building a chat or agent workload on: Google’s Lyria 3 Clip Preview and Lyria 3 Pro Preview are audio/music generation models, and OpenRouter’s own Free Models Router is a routing utility that dispatches to whichever free model is available, not a distinct model with its own weights. Count those in and you get the full 25; count only models you’d actually target directly by name and it’s 22.
What stands out
Biggest context windows, free: NVIDIA’s Nemotron 3 Super and Nemotron 3 Ultra both carry 1,000,000-token context windows at zero cost — larger than most paid mid-tier models. Qwen3 Coder 480B A35B matches that ceiling at 1,048,576 tokens and is a genuinely large model (480B parameters, 35B active) to get for free.
Real production names, not just demos: Llama 3.3 70B and gpt-oss-120b are both models teams actually run in production paid tiers elsewhere. Getting them free isn’t a stripped-down trial version — it’s the same weights.
Smallest, fastest: Liquid’s LFM2.5-1.2B models (Instruct and Thinking variants) and Meta’s Llama 3.2 3B sit at the other end — small enough to be genuinely fast for latency-sensitive, low-complexity tasks, with 32,768 and 131,072-token windows respectively.
Reasoning variant included: LFM2.5-1.2B-Thinking is one of the few free models on the list built for chain-of-thought-style reasoning rather than direct instruction-following, worth trying if you want reasoning behavior without paying for it first.
The honest caveat: free tiers rotate
Free-model availability on aggregators is not stable, and this isn’t unique to us. A documented complaint about OpenRouter’s free tier is that its roughly 29 rotating free models change without much warning — a model that’s free this month may be metered or removed next month, which quietly breaks any benchmark, demo, or low-stakes production path built assuming it stays free.
We’re not going to pretend that risk doesn’t exist here too. Providers set their own free-tier terms, not us, and a lab can pull or reprice a free offering at any time. What we commit to is not hiding it when it happens.
Our commitment: a dated availability log
Every addition, removal, or repricing of a free model on OpenKey gets a dated entry in the changelog. If a model drops off this list, you’ll be able to see exactly when and cross-reference it against your own usage instead of discovering it as a silent 402 in production. That’s a small thing, but it’s the difference between “the free tier changed” being a debugging mystery and being a two-second lookup.
If you’re building on a free model, treat it as free-for-now: pin the model ID in your code (not a -latest alias), watch the changelog, and have a paid fallback path ready. For workloads where “free” needs to mean “free and stable,” the safer bet is one of the sub-$0.50/M input workhorse models — cheap enough that rotation risk stops mattering.
Full live list with pricing badges: /models/free. Full catalog: /models. Price comparison across everything: /pricing.