Kimi K3.1 Pro: Moonshot's 3T Open Model Chases Claude Opus

Kimi K3.1 Pro is Moonshot AI's first big upgrade since Kimi K3, and it landed on Friday, October 23 with a loud claim: near-frontier coding at a much lower price. Moonshot says the 3T-class model scores 65.7% on Terminal-Bench 4.0, a hair behind Claude Opus 5.5, and the full weights are promised by November 4. Polymarket already has a market on how it will debut on Humanity's Last Exam.

What actually happened

Moonshot's official blog calls the release "our most capable 3T-class open-weight model to date." It is a post-trained version of K3 with much more reinforcement learning compute, aimed at coding, 3D creation, financial research and agent tasks. A Highspeed version runs two to three times faster, Moonshot says.

The model is live now on Kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the Kimi API. The weights are not public yet. Moonshot says they will be released by November 4, 2026, so "open" is a promise for now, not a download.

Bloomberg framed it as the first major upgrade from a startup now valued around $50 billion, after what it called Moonshot's own DeepSeek moment over the summer.

The benchmark claims, with asterisks

Every number here is self-reported by Moonshot. The Next Web says K3.1 Pro scored 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. MarkTechPost adds that it leads SWE-Marathon at 51.7, but trails badly on Terminal-Bench Science, scoring 38.1 versus 68.1 for GPT-6 Astra.

Moonshot also claims fewer made-up facts. On its internal hallucination stress test, the share of false claims fell to 8.3% from 23.4% for K3. MarkTechPost notes GPT-6 Astra still scores lower on that same test.

Price is the real flex. MarkTechPost lists $2.00 per million input tokens and $10.00 per million output, down from $3 and $15 for K3, with the Highspeed variant at $4 and $20. API access unlocks after a minimum $1 top-up, according to Moonshot's quickstart docs.

The cyber angle

The most unusual part is safety. The Next Web reports that Moonshot removed hacking-related material from training so the open release would not push the limits of free models. An unfiltered checkpoint solved 90.5% of Moonshot's in-house exploit tasks. The shipped model solved 81%, still far above K3's 33.3%.

That matters because open weights cannot be recalled once downloaded. Washington is also watching: The Next Web notes the White House accused Moonshot in July of using banned Nvidia chips and copying an Anthropic model to build K3, without publishing evidence.

What Polymarket thinks

Polymarket's market "Kimi K3.1 Pro: HLE-Diamond Debut?" resolves on the score shown on the Humanity's Last Exam site at 12:00 PM ET the day after the model first appears there. Our final snapshot from the Polymarket API at 06:21 UTC on October 24:

• 20% or higher: about 88.5%
• 25% or higher: about 74.5%
• 30% or higher: about 39% (bids 37, asks 41 cents)
• 35% or higher: about 16%

The headline question, "Will Kimi K3.1 Pro debut with an HLE-Diamond score of 30% or higher?", sat near 39% Yes, down from about 44% in our first pull earlier that morning. Volume was small, about $3,700 across the event, so a single trader can move these prices.

Read it as: traders expect a strong debut, but they see 30% as somewhat less likely than not and 35% as a long shot.

What to do as a reader

If you build with models, test it on your own tasks before switching. Self-reported scores are a starting point, and the science gap shows the model is not best at everything. For context on Moonshot's last release, see our Kimi K2.6 piece, and for the open-model price war, our DeepSeek V4.1 Flash 1017 review.

If you trade the HLE market, read the rules on Polymarket: only the first listing counts, and thin books move fast. Do not treat any AI hype as a reason to ape into tokens that borrow the Kimi name. Nothing here is investment advice.

Sources: Kimi blog, The Next Web, MarkTechPost and the Polymarket market.