Kimi K3.1 Pro Scores 65.7% on Terminal-Bench 4.0, 0.7 Points Behind Opus 5.5, and the Weights Land by November 4. Here's What to Test Before You Care.
Moonshot released Kimi K3.1 Pro and a faster K3.1 Pro Highspeed variant, and its own blog reports 65.7% on Terminal-Bench 4.0 against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Full open weights are promised by November 4. A 0.7-point gap to the best closed coding model, from a model you may be able to run yourself, is exactly the kind of headline that makes people rewrite their stack on a Tuesday.
What Moonshot actually claims
All the numbers below come from Moonshot's own announcement, so treat them as vendor claims until someone else reproduces them. The model has a 1 million token context window. The Highspeed variant is described as 2 to 3 times faster than K3.1 Pro, and its API ID is kimi-k3.1-pro-highspeed.
On Terminal-Bench 4.0, K3.1 Pro scores 65.7%. On BenchCAD with tools it scores 92.7%, behind Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. Moonshot also reports a claim-level hallucination rate on its own stress test that fell from 23.4% for K3 to 8.3%, and it notes that this is a stress test, not a typical error rate. I like that caveat. It is rare to see one on a launch page.
Why I would not trust the headline benchmark
On October 24 I wrote about an audit of Terminal-Bench 4 by Preference Model, which reviewed 65 tasks and found 60 with at least one issue, 44 that could be passed without doing the work and 41 that were unfair to honest solutions. A model that scores 65.7% on a benchmark like that is telling you something, but I cannot say what.
That does not make Kimi's number false. It makes the 0.7-point gap meaningless. Differences that small on a flawed test are noise, and the useful reading is the loose one: this model is in the same neighborhood as the leaders on agentic terminal tasks. That is worth a test. It is not worth a migration.
What matters to a solo operator
Open weights are only useful to you if three things line up. First, the license. Moonshot calls the model open-weight, but the announcement does not state license terms, and a license that restricts commercial use or large deployments changes everything. Second, the price someone will charge you to host it, since almost nobody running a one-person product wants to rent a GPU cluster. Third, whether it holds up on your own tasks.
On the second point, a competitive open model does a useful thing even if you never use it: it gives you leverage. The price cuts across closed providers this autumn are partly a response to substitutes being easy to swap in. A strong open model on a dozen hosting providers keeps that pressure on.
A 30-minute test for November 4
When the weights drop, do this before you form an opinion. Read the license first, because five minutes there can end the whole exercise. Then take five real tasks from your own repo, such as a failing test, a small refactor and a dependency bump, and run each through your current agent and through Kimi on a hosted endpoint. Record whether it finished, how long it took and what it cost. Do not write down a score. Write down which result you would have accepted without editing.
The Highspeed variant deserves its own run if you do interactive work, since latency is where a lot of coding agents lose people.
Where this could be wrong
Moonshot's other results look strong, and Terminal-Bench is only one measure. It is possible that K3.1 Pro is simply very good, and that my skepticism about the benchmark is hiding a real jump. I also have not seen independent runs, so the opposite is just as possible. That is why the test is on your tasks, not on a leaderboard.
What I'd actually do
Nothing until November 4, then the test above, license first. If it passes and a hosting provider offers it at a clear discount, route your lowest-risk agent tasks to it behind a model switch you can flip back. Do not rebuild anything around one launch post.
Author
Lukas
@lukcombinator