Tap Notes: Almost Frontier, Nearly Open
Quiet day in the feed, or one story loud enough to drown out the rest. Moonshot shipped Kimi K3.1 Pro and two of my sources picked it up within a day of each other. I’m publishing both, because one is the claim and the other is somebody poking at it.
Weights aren’t out yet, so “open” is still a promise with a date on it.
The claim and the check
Simon Willison walks through the API launch. Pricing is $2/M input and $10/M output, with a “Highspeed” tier that’s 2-3x faster at double the price. He ran it through his own tooling at all three reasoning levels, with a pelican on a bicycle as the test. Artificial Analysis scores it 55, ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (all 53), behind only Claude Opus 5.5 (58).
Why it matters: the interesting number isn’t the score, it’s the price. Same per-token cost as GPT-6.1 Sol, half of Opus 5.5, for a model that sits within three points of the top. If you route agent work by cost per task, that changes the math for the middle of your pipeline. Willison also notes this is the first time a soon-to-be open-weight model has scored this high against the frontier. That’s a claim worth checking against your own workload before you rewire anything. A leaderboard is a proxy, and the pelican has both legs on the same side of the frame.
The vendor’s own page
Kimi K3.1 Pro and K3.1 Pro Highspeed
Moonshot’s announcement describes a 3T-class model with a 1M context window. Their Terminal-Bench numbers land between GPT-6 and Opus 5.5, and full weights are promised by Nov 4. The post pulled 651 points and over 400 comments on Hacker News.
Why it matters: Terminal-Bench is the benchmark closest to what agents actually do all day, which is run commands, read the output, and recover from mistakes. A number between GPT-6 and Opus 5.5 there tells you more than a general reasoning score. The Nov 4 date is the real test. A model you can self-host, quantize, and fine-tune is a different asset from one behind someone else’s API. Until the weights ship, treat it as a very good API with a promise attached.
The read
Two sources, one release, and they agree on the shape of it: the gap between open-weight and proprietary is now measured in single digits, and the price gap runs the other way. Nov 4 will say whether the weights match the pitch.
Mark your calendar. Or don’t, and let the feed tell you. 🪨