Chinese AI lab Moonshot AI has shipped Kimi K3.1 Pro, and the leaderboard just got uncomfortable for everyone else. The new model lands at 55 points on the Artificial Analysis Intelligence Index — fourth place overall, three points behind Claude Opus 5.5 and two behind Claude Sonnet 5.5. For a model that's about to be handed out as open weights, that's a serious statement.
What happened
Kimi K3.1 Pro isn't a new architecture. It's the same Kimi K3 with the same 2.8 trillion parameters, but Moonshot spent a lot more compute on post-training — the stage where a model gets drilled on how to reason, follow instructions and use tools properly, instead of just predicting the next word. Same engine, much better tuning.
The payoff is an 11-point jump on the Artificial Analysis Intelligence Index over K3. Only three models now rank higher: Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56. Alongside the standard model, Moonshot is also shipping K3.1 Pro Highspeed, a variant it says runs two to three times faster.
The open-weights deadline that matters
Moonshot plans to release the model weights by November 4, 2026. If that happens on schedule, K3.1 Pro becomes the highest-scoring open-weights model anywhere — and the gap between the best downloadable model and the best rented one shrinks from 12 points to 3.
"Open weights" means you can download the model itself and run it on your own machines, with no per-token bill and no data leaving your network. Proprietary models like Claude you can only rent through an app or an API. Three points is close enough that a lot of teams will happily take that trade.
Agents are where it really moved
Agentic tasks — where the model plans several steps, picks up tools and finishes a job without a human nudging it along every move — showed the biggest gains.
On GDPval-AA it reaches 1,846 Elo, a 312-point jump over K3, trailing only Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). On Terminal-Bench 4.0, Moonshot reports 65.7%, ahead of GPT-6 Astra at 58.2% and a whisker behind Claude Opus 5.5 at 66.4%. K3 managed just 13% on the same test — that's the difference between a model that talks about a task and one that actually finishes it.
Fewer invented facts, better spreadsheets
On Moonshot's internal Kimi Hallu Bench Hard stress test, claim-level hallucinations dropped from 23.4% to 8.3%. That's still behind GPT-6 Astra's 5.5% on the same test, and Moonshot is upfront that it's a stress test rather than a measure of everyday error rates. Still, going from roughly one bad claim in four to about one in twelve is the kind of change people feel in daily use.
On internal finance benchmarks the model gained 7.2 points on Financial Modeling and 11.6 points on Financial Research, and it tops every model Moonshot tested on OfficeQA Pro v2 for general office work. On 3D work it plugs straight into Unity, Unreal Engine and Godot, handling everything from quick fixes to full project builds.
Same size, smaller bill
Parameter count stays flat at 2.8 trillion, but API prices fall 33%. You'll pay $2.00 per million input tokens, $10.00 per million output tokens and $0.20 per million cached input tokens. A token is roughly three-quarters of an English word, so a million tokens goes a long way — you can push a whole stack of documents through the model for roughly the price of a coffee.
K3.1 Pro Highspeed costs $4.00 for input and $20.00 for output per million tokens, with the same $0.20 cache rate. Both versions are live now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API.
What it means for you
You don't need to care about Elo ratings to feel this release. Models in the 55-point range stop falling apart halfway through a plan, which changes what you can realistically hand off.
At home
Planning is the easy win. A model that holds constraints can build a two-week meal plan around a grocery budget, draft a packing list for a family trip, or turn a pile of school emails into one clear calendar. The old frustration — asking for twelve steps and getting nine, plus two invented ones — is exactly what the hallucination and agentic gains chip away at.
At work
Anything with a paper trail is fair game: summarizing a long thread, turning meeting notes into an action list, checking a contract clause against your own policy. Terminal-Bench is a coding test, but the skill it measures — sticking with a multi-step job until it's done — is the same skill that makes a model useful for chasing down a spreadsheet error at 5pm on a Friday.
Running a small business
Here the pricing matters as much as the score. A 33% cut at $2.00 input and $10.00 output per million tokens means you can put a model behind customer emails, product descriptions and invoice checks without the bill eating your margin. If you want to compare it against other free options before you commit, mykreatool is a handy place to test free AI tools side by side in one tab.
Studying
K3.1 Pro tops Moonshot's own OfficeQA Pro v2 benchmark, and document work is where most students actually need help: pulling themes out of a 60-page reading, building practice questions from lecture notes, checking your own argument for gaps. Lower hallucination rates matter most here, because a confidently wrong fact is worse than no answer at all.
Creative work
On BenchCAD, a CAD benchmark, it lands right at the front of the pack once tools are switched on, and the model works inside Unity, Unreal Engine and Godot. If you make games, levels or 3D assets, that's a real workflow change — describing a scene and getting working project code back beats copying snippets between tools. Writers and designers get the usual benefit: fast first drafts you then rewrite in your own voice.
Earning extra income
Cheap, capable models make small automation services viable. Think replying to leads for a local plumber, generating listing descriptions for a reseller, or building a research brief service for freelancers. At $0.20 per million cached tokens, repeat work that reuses the same instructions gets very inexpensive very fast.
How to try it right now
Start free, then scale up only if it earns its keep.
1. Open Kimi.ai. Sign in and use the free chat. No card, no API key, no setup.
2. Give it a real job, not a riddle. Paste an actual document, email thread or spreadsheet extract you're already working on. Toy prompts tell you nothing.
3. Run the same prompt through whatever you pay for today. Same input, same instructions. Compare the output, not the vibes.
4. Try Kimi Code if you write software. It's the coding-focused surface and the natural home for Terminal-Bench-style work.
5. Use Kimi Work for team documents. Good for meeting notes, specs and internal write-ups.
6. Get an API key only when you've found a repeatable task. Start on the standard model at $2.00 input and $10.00 output per million tokens. Switch to Highspeed ($4.00/$20.00) only if waiting on responses is actually costing you something.
7. Set a reminder for November 4, 2026. If the weights drop as promised, you'll be able to run the model yourself — worth knowing before you sign an annual contract anywhere.
Upsides and what changes
The headline is the price-to-performance ratio. You're getting near-frontier scores at a third less than before, on a model Moonshot intends to make downloadable. On SWE Marathon v1.1 and DeepSWE v1.1 — two long-horizon software engineering tests — it edges past Claude Opus 5.5 outright. Those aren't benchmark quirks, either — that's a genuine competitor.
The bigger shift is structural. Once the best open-weights model is within three points of the best proprietary one, "we can't use AI because of cost or data rules" gets a lot harder to defend. Companies with privacy constraints, teams in regions with patchy API access, and anyone building on top of a model rather than renting one all gain real options.
Limitations
Go in with open eyes. A lot of these numbers are vendor-reported — Terminal-Bench, the finance benchmarks and OfficeQA Pro v2 come from Moonshot itself, and every result in its table was run with reasoning effort set to "max," which is the slowest and most expensive setting, not what you'll get by default. The Artificial Analysis Index is a single composite score, so fourth place doesn't mean fourth in everything. ExploitBench is the clearest counterexample: on that security-focused test it trails both Claude Opus 5.5 and GPT-6 Astra by a wide margin, so if security and exploit analysis is your use case, this model isn't the leader. The open weights are a plan, not a shipped file, and deadlines slip. And all of this runs through a Chinese provider, so check your own data policies before you paste anything sensitive into a hosted chat.
Conclusion
The gap between the best AI you can rent and the best AI you can download has narrowed to three points, and the price came down 33% along the way. That's the whole story in one sentence.
Here's your move for today: open Kimi.ai, paste in the single most annoying recurring task you did last week, and see whether the free version handles it. Fifteen minutes, no card, and you'll know whether this belongs in your workflow before the open weights even land.
Comments 0