Kimi K3.1 Pro is the same base as K3 with far more RL, and it trails Opus 5.5 by under a point on Terminal-Bench

On 23 October Moonshot AI shipped Kimi K3.1 Pro to its API and apps. The base model stayed the same: the company itself says K3.1 Pro is a post-trained version of K3 with substantially larger post-training compute.

Moonshot kept the K3 base and poured in far more reinforcement learning

To run joint RL across hundreds of domains, Moonshot says it rebuilt the whole stack: the RL framework, long-context inference, rollout scheduling and sandboxes. The result is a model with a 1-million-token context window that matches or approaches the strongest proprietary models on coding, agentic and knowledge tests. On Terminal-Bench 4.0 it reports 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On its own Kimi Code Bench 3.0 Long-horizon it climbs from 55.95% for K3 to 61.2%.

Nathan Lambert, who previously co-led the Olmo project at Ai2 and wrote a book on post-training, summed it up on X in a single post. The original K3 release, in his reading, was a very impressive base model with lighter post-training, and now it looks like Moonshot has figured out its RL. China, he adds, is very much keeping pace with capability gains.

Independent measurement by Artificial Analysis puts the model at 55 on its Intelligence Index. Simon Willison notes that this is ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53, and behind only Claude Opus 5.5 at 58. The API costs $2 per million input tokens and $10 per million output tokens, the same as GPT-6.1 Sol and half the price of Opus 5.5. The weights are not out yet; Moonshot promises them by 4 November 2026.

Post-training is becoming the faster road to the frontier

For the market this matters more than any single score. If the same base can jump this far with more RL, the gap between Chinese and American labs is no longer measured only by who has more chips for pretraining. Post-training can be iterated more often and in smaller batches, and that is where Moonshot spent its compute. For US labs leaning on chip export controls, it is an uncomfortable signal: hardware limits slow down building a new base, but they do much less to stop catching up through RL.

For developer teams, model selection shifts. A model priced like GPT-6.1 Sol and scoring just below Opus 5.5 is due to arrive as open weights within days. Teams that cannot send code or documents to a Chinese cloud get an option that does not require Moonshot's API. It will, however, be a 3T-class model, so only teams with a GPU cluster will be able to run it themselves.

Moonshot measured most of these numbers itself

Terminal-Bench was run in the Claude Code harness, Kimi Code Bench is an internal test, and competitor scores are taken from their own blogs and leaderboards. In an internal stress test, claim-level hallucinations fell from 23.4% to 8.3%, and Moonshot itself adds that these are not everyday error rates. Artificial Analysis also recorded that the model produced 150 million output tokens while running the index, against a median of 87 million. The price per million tokens therefore says little about the price per task: a verbose model gets billed many times over.

The second layer concerns cybersecurity. Moonshot says it removed cyber-related data from training. A version without those interventions solved 90.5% of tasks on the internal xdev-bench-hard test; the final K3.1 Pro solved 81.0%. The model's offensive capability is therefore deliberately lower than what the pipeline can produce, yet still far above K3. Where that line goes next is, for now, the lab's call alone.

4 November and the first independent weight tests will decide it

The first signal is simple: will Moonshot publish full weights by 4 November, and under what licence? The second will come from independent teams running the model in their own harnesses to check whether the Terminal-Bench result holds without Claude Code. The third will show whether Lambert's thesis holds more broadly. If DeepSeek or Qwen show a similar jump from post-training alone, RL becomes the main racetrack for Chinese labs.

Lilith's verdict

Moonshot took the same engine and simply ran it in on the track for much longer. Anyone in Washington who believed guarding the engine factories was enough is now learning that the race is decided in the pit lane.