TL;DR: Moonshot AI just put Kimi K3.1 Pro live across its apps and API, and the numbers are the story. On Terminal-Bench 4.0 it scores 65.7%, ahead of OpenAI’s GPT-6 Astra at 58.2% and just a step behind Anthropic’s Claude Opus 5.5 at 66.4%. It’s API-only today, but Moonshot says the full weights ship by November 4, which would make this the strongest open-weight model anyone has released to date.
Moonshot AI pushed Kimi K3.1 Pro live today across Kimi.ai, Kimi Work, Kimi Code and the Kimi API. It’s the first model in a new K3.1 series, built as a post-trained version of Kimi K3, the lab’s 3-trillion-parameter-class open-weight model that turned heads earlier this year. Moonshot describes K3.1 Pro as “our most capable 3T-class open-weight model to date,” and the gap between this release and its predecessor is what makes it worth paying attention to.
The headline number is the context window: 1 million tokens, enough to hold an entire codebase or a stack of lengthy documents in a single pass. Moonshot shipped it in two flavors. K3.1 Pro is the standard model, and K3.1 Pro Highspeed runs two to three times faster for roughly double the price, aside from cache-hit input pricing. Both are priced at $2 per million input tokens and $10 per million output tokens on the Kimi API, with three selectable reasoning levels: low, medium and max.
What actually changed between K3 and K3.1 Pro is substantially larger post-training compute, not a bigger base model. Moonshot’s own benchmark table backs that up. On its internal Kimi Code Bench 3.0 long-horizon test, the new model jumps to 61.2% from K3’s 55.95%. On financial-reasoning tasks the gains are bigger still: Business Operation scores climb from 8.9 to 37, and Financial Research moves from 58.8 to 70.4. Moonshot also says its claim-level hallucination rate on an internal hard benchmark dropped from 23.4% with K3 down to 8.3% with K3.1 Pro, which is the kind of number that matters more to people actually deploying these models than raw intelligence scores do.
Independent AI writer Simon Willison, who tracks model releases closely and has no stake in Moonshot’s success, called this “the first time a (soon-to-be) open weight model has scored this highly against the frontier.” That’s a meaningful statement coming from someone who has watched a long line of “open-weight challenger” announcements come and go without closing the gap. On Moonshot’s BenchCAD test with tool use enabled, K3.1 Pro posts 92.7% against Claude Opus 5.5’s 96.2% and GPT-6 Astra’s 95.9%, a real gap but not the kind of gap that used to separate Chinese open-weight labs from the US frontier.
There’s a catch worth sitting with before anyone celebrates how open this actually is. The weights aren’t public yet. Moonshot says the full release lands by November 4, which is less than two weeks out, but until then K3.1 Pro is only reachable through Moonshot’s own hosted API and apps, same as any closed model from OpenAI or Anthropic. Calling something “open-weight” before the weights exist is a label the whole industry has gotten comfortable stretching, and Moonshot is doing the same thing here even as it builds its reputation on openness.
The more interesting disclosure is what Moonshot says it deliberately left out. The company confirmed it stripped cyber-related data from both training stages specifically so the open release wouldn’t “extend the absolute frontier of cyber capability.” On its internal cyber benchmark, xdev-bench-hard, the public-facing checkpoint scores 81% pass rate, down from 90.5% on the uncensored version Moonshot tested internally. That’s a notable admission: the lab is openly saying it built a more capable model than the one it plans to hand out, and trimmed the difference on purpose rather than quietly shipping whatever scored highest. It’s the same tension that showed up when DeepSeek open-sourced its DeepInfer inference engine earlier this month: Chinese labs keep pushing real capability into the open ecosystem, and keep drawing their own lines about exactly how much of it to hand over.
The timing matters for a different reason too. Every time an open-weight model closes in on Opus-class or GPT-6-class performance, it resets what “frontier” means for teams that can’t or won’t pay API rates from the big three. A 3-trillion-parameter model is still enormous and won’t run on a laptop, but once the weights are out, any company with serious GPU budget gets a model that trades blows with Claude Opus 5.5 on agentic coding tasks without a per-token bill to Anthropic. That’s the same dynamic we wrote about when Hugging Face released its open Carbon-X model: open weights don’t just save money, they change who gets to build on top of frontier-level AI at all.
Enterprise users should also note that K3.1 Pro is off by default. Admins have to turn it on for their organization before anyone can use it, which is a small detail but a sensible one given how quickly a new reasoning model can change outputs an enterprise team has already tuned workflows around.
None of Moonshot’s numbers here come from a neutral third party. These are vendor-reported benchmarks, run under Moonshot’s own test conditions, and Simon Willison’s cross-reference to the Artificial Analysis Intelligence Index (where he puts Kimi around 55 against Opus 5.5’s 58) is itself secondhand. Treat the specific percentages as Moonshot’s best self-portrait rather than an audited result, the same caution worth applying to any lab’s launch-day benchmark table.
The bottom line
Kimi K3.1 Pro is the clearest sign yet that the gap between frontier closed models and open-weight Chinese models is shrinking faster than most people outside the AI research community have noticed. A model that beats GPT-6 Astra and nearly matches Claude Opus 5.5 on real benchmarks, with weights due in under two weeks, is not a research curiosity. It’s a genuine option for any team deciding whether to keep paying frontier-lab API prices. The real test comes November 4, when we find out whether the public weights perform like the benchmarks promise, or whether the gap Moonshot left in place by stripping cyber data shows up elsewhere too.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
