Kimi K3.1 Pro: open weights Nov 4, read the whole table
Moonshot's Kimi K3.1 Pro beats GPT-6 Astra on Terminal-Bench but trails Claude Opus 5.5. Open weights land by Nov 4. What to test before you switch.
Kimi K3.1 Pro is out, and the headline number is not the whole story. Moonshot AI's K3.1 Pro announcement puts the model at 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2%. The same table shows Claude Opus 5.5 at 66.4%, still on top. Kimi K3.1 Pro is available now on the Kimi API, and Moonshot says the full model weights will be released by November 4.
What actually happened
From Moonshot's own post:
- Two API models.
kimi-k3.1-proandkimi-k3.1-pro-highspeed, the second described as 2 to 3 times faster. - 1 million token context window.
- Zero Data Retention is available for API customers.
- Open weights by November 4. The post does not name a license yet.
- Coding numbers. Terminal-Bench 4.0 at 65.7%, and BenchCAD with tools at 92.7%, behind Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
- Big internal gains over K3. Long-horizon coding, finance tasks and a hallucination stress test where claim-level errors fell from 23.4% to 8.3%. Those are Moonshot's internal benchmarks; nobody outside has reproduced them yet.
We did not find per-token pricing for K3.1 Pro on Kimi's public pricing pages at the time of writing, so we are not quoting one.
Why it matters for your business
Every benchmark here is vendor-reported. "Tops GPT-6 Astra" is accurate for one test and wrong as a general claim. Read the full table, not the press line, and run your own evals on your own repo before you move a workflow.
Open weights change the options, not the cost. Weights by November 4 mean you could host it yourself or buy it from a third-party host. As we wrote when K3 opened up, a model of this size is not cheap to run, and the license will decide what you can do with it commercially. Wait for the license text.
Zero Data Retention is the line to check first. If your code or customer data goes to the API, confirm ZDR is switched on for your account before the first call, not after.
This is why your model layer should be swappable. A new contender every few weeks is the normal state of things now. If your agents call one provider directly, every launch like this is a rewrite. If they go through a router, it is a config change and an eval run.
Key takeaways
- Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0, above GPT-6 Astra, below Claude Opus 5.5
- Available now as kimi-k3.1-pro and a 2-3x faster highspeed variant, with a 1M token context
- Full weights promised by November 4; the license is not named yet
- Most gains are on Moonshot's internal benchmarks, so test on your own code
- No public per-token price found yet; confirm cost and ZDR before routing traffic
Make new models a config change, not a rewrite. We build vendor-agnostic agent stacks with a model router and an eval suite on your real tasks, so you can try K3.1 Pro on Monday and keep it only if it wins. See how we build, or tell us which models you run today.
Sources: Moonshot AI: Kimi K3.1 Pro, Kimi API pricing.
- #kimi
- #open-weights
- #coding-models
- #benchmarks
- #model-routing
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Plex2Shell: update Plex Media Server, then check for miners
Plex2Shell chains four Plex Media Server flaws into remote code execution. CISA lists it as exploited. Update to 1.43.3.10861 and check the box for implants.
Read itCopilot usage metrics API now names your MCP servers and agents
GitHub's Copilot usage metrics API now reports MCP servers, custom agents and skills by name, not as 'other'. Use it to audit what your AI coding fleet runs.
Read it