Free Coding Models on OpenRouter (August 2026): Ox Alpha and 7 More Worth Your Requests
Ox-alpha headlines OpenRouter's free collection right now, but it's not alone — Kimi K2.6, gpt-oss-120b, Nemotron 3 Ultra and more. Plus the rate limits (20/min, 50/day) that actually decide what you can build on $0.
Priya Anand · Technical Writer · August 26, 2026 · 6 min readShort answer: yes, you can run a real coding agent on $0 right now. Ox-alpha — the anonymous stealth model with a 1M context window — is the headliner, and it's joined by a deep bench of open-weight models. The catch isn't the models; it's the rate limits. Here's the full field, and how to actually build with it.
The free field, August 2026
| Model | Context | Best for |
|---|---|---|
| Ox Alpha (stealth) | 1M | Reasoning, long-horizon agentic coding — free during preview |
| Kimi K2.6 (Moonshot AI) | 262K | Strongest free open-weight agentic coder (47.6% on CursorBench 3.1) |
| NVIDIA Nemotron 3 Ultra | 1M | Long agentic workflows; hybrid Mamba-Transformer MoE |
| Poolside Laguna S 2.1 / M.1 | 262K | Terminal/agentic coding; 70.2% Terminal-Bench 2.1; open-weight |
| Cohere North Mini Code | 256K | Agent-harness tasks; 30B MoE, Apache 2.0 |
| OpenAI gpt-oss-120b | 131K | Dependable open-weight generalist |
| MiniMax M3 | — | Coding + tool use, high/xhigh reasoning effort |
One absence worth noting: Qwen3 Coder 480B (1M context), the long-time free favorite, was flagged for retirement around July 19 — a reminder that this landscape reshuffles without warning.
Connect the Claude or Codex you already pay for — the rest runs on workers that cost a fraction.
Download meshcode →The rate limits that actually decide things
Free on OpenRouter doesn't mean unlimited. The published free-tier limits:
- 20 requests/minute on
:freemodels - 50 requests/day on an account with no credits
- 1,000 requests/day once you've bought $10+ in credits (one-time, and the credits are yours to spend)
Two levers most people miss:
openrouter/freeis an auto-router: point your agent at it and it picks any free model that satisfies the request's needs (tool calling, image input, structured output).:freeappended to a model ID pins that model's free endpoint, e.g. for Kimi K2.6.
Do the math for agent work: 50 requests a day disappears in about ten minutes of a chatty agent loop. At 1,000/day, you're budgeting again — just at a very different price point.
How to build on $0 without fighting the meter
The teams getting real work out of free models share the same shape:
- Fewer, longer sessions. A 1M-context model like ox-alpha or Nemotron 3 Ultra can hold an entire working session in one conversation. Every context re-read you avoid is a request you save.
- Parallel panes, not serial loops. Split work across agents — one per service, per repo, per task — instead of one agent making hundreds of calls. Four panes at 20 req/min each beats one agent thrashing against the cap.
- Save the premium model for the hard 10%. Free models handle the bulk; route the gnarly debugging to a frontier model when it's actually warranted.
Running them in MeshCode
MeshCode's OpenRouter integration puts all of the above behind one toggle: Settings → Models → OpenRouter, paste your key, and pick models from a 418-model picker — ox-alpha included. Each pane runs its own agent and its own model, so you can put ox-alpha on three panes, Kimi K2.6 on the fourth, and keep Claude for the parts that need it. Full walkthrough in Ox Alpha: the free stealth coding model.
The honest caveat: free endpoints are a moving floor. Models get retired, previews end, latency spikes at peak. Treat $0 as a way to test more, fan out wider, and learn faster — not as a production architecture.
Free-model availability and limits per OpenRouter's published docs as of August 26, 2026, and subject to change. MeshCode is not affiliated with OpenRouter or the model providers listed.
More from the blog
Ox Alpha: The Free Stealth Coding Model (and How to Run It in MeshCode)
A frontier-class model appeared on OpenRouter under the name ox-alpha — no lab, no paper, $0 per token, 1M context. Here's what's actually known about it, and how to put it to work in MeshCode today.
Priya Anand
Ox Alpha Benchmarks: What the Numbers Actually Say
The viral '80% on DeepSWE — beats GPT-5.6 Sol' claim came from a 10-task sample. Full 113-task runs measured 58.4%. Here's every number, the tool-call failure caveat, and what it means for agent work.
Priya Anand
Who Made Ox Alpha? Every Clue, From Tokenizer Fingerprints to Pony Alpha
No lab has claimed ox-alpha, the free stealth coding model on OpenRouter. We rank every attribution clue — the 44/44 tokenizer match to GLM-5.3, the compression analysis, the self-identification leak, and the theories that don't hold up.
Yuki Tanaka