AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Chinese AI model Kimi K3.1 Pro is closing in on Anthropic's Claude Opus 5.5

Collected Oct 23, 2026

Moonshot AI has released Kimi K3.1 Pro, a post-trained version of its Kimi K3 model trained with substantially more post-training compute. On the Artificial Analysis Intelligence Index it scores 55 points, an 11-point gain over K3 that places it fourth overall. The three models ahead of it are Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56. Moonshot is also shipping a variant called K3.1 Pro Highspeed that it says runs two to three times faster. Both are available now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API.

Moonshot plans to release the model weights by November 4, 2026. If that happens, and if nothing else changes in the meantime, K3.1 Pro would become the highest-scoring open-weights model available and would narrow the gap between open and proprietary models from 12 points to 3. That framing matters: the benchmark is a third-party index maintained by Artificial Analysis, and a narrow top-of-table gap does not mean parity across every task. It does suggest that openly available weights are tracking the proprietary frontier much more closely than they did a generation ago, which is the kind of shift that changes build-versus-buy calculations for teams choosing a base model.

Agentic work is where the gains are largest. On GDPval-AA, which Artificial Analysis reports, K3.1 Pro reaches 1,846 Elo, 312 points above K3. That puts it just behind Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra at 58.2 percent and narrowly behind Claude Opus 5.5 at 66.4 percent. K3 managed only 13 percent on the same test, so the jump there is the most dramatic number in the release. Terminal-Bench targets multi-step command-line work, the kind of loop an agent runs when it plans, executes and corrects itself inside a shell, so a score near the top proprietary models is a meaningful signal for tool-using pipelines.

Hallucination rates fell sharply on Moonshot's own evaluation. On its internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent with K3 to 8.3 percent. Moonshot describes this as a stress test rather than a measure of general error rates, which is the right caveat: internal benchmarks are tuned to the areas a lab cares about and are not directly comparable across vendors. GPT-6 Astra records 5.5 percent on the same test, so it retains an edge there. On the company's internal finance benchmarks, K3.1 Pro gained 7.2 points on Financial Modeling and 11.6 points on Financial Research, and Moonshot says it tops all tested models on OfficeQA Pro v2 for general office tasks.

The model also works directly with game engines including Unity, Unreal Engine and Godot, handling quick fixes and full project builds, according to Moonshot. On BenchCAD it scores 92.7 percent with tools, close to Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent; without tools it reaches 72.8 percent against 73.0 percent for Claude Opus 5.5. Moonshot says it is strongest at parametric modeling of mechanical parts, architectural models and configurable products, and that it can write and run scripts for Blender and CAD software without MCP connectors. MCP here refers to the Model Context Protocol, the common interface many agents use to reach external tools; skipping it means the model can drive those applications more directly rather than through an integration layer. Moonshot also launched a feature called 3D Make on Kimi.com that lets users preview, edit and export 3D models.

The architecture is unchanged in size. K3.1 Pro keeps 2.8 trillion total parameters with 104 billion active, a sparse mixture-of-experts design where only a fraction of the network runs per token, which is how a model of that scale stays servable. The context window spans one million tokens and the model handles both text and images. Artificial Analysis places it at the top of the Pareto frontier, meaning no model with a comparable parameter count scores higher on the index.

Pricing improves. Token prices drop by 33 percent: input costs $2 per million tokens, down from $3; output costs $10, down from $15; and cache hits cost $0.20, down from $0.30. The cost per Intelligence Index task is $1.98 versus $2.00 for K3, almost unchanged, because the model generates 14 percent more output tokens. In practice this is a familiar trade-off: cheaper unit pricing can be partly absorbed by more verbose reasoning, so teams budgeting agents should measure cost per completed task rather than per token, and per-task cost is exactly the metric where the two models look flat.

Moonshot devotes a full section of its technical write-up to cybersecurity. It says vulnerability research capabilities kept growing as post-training scaled, because cyber-related content sat in the general mid-training corpus and related tasks were part of the reinforcement learning environment mix. For the open-weights release, Moonshot stripped cyber-related data from both training phases. On its internal xdev-bench-hard, which reproduces recently disclosed vulnerabilities in software such as the Linux kernel, Node.js and JavaScriptCore, the uncleaned version solves 90.5 percent of tasks; after cleaning, K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash reaches 85.7 percent on the same test and Kimi K3 scored 33.3 percent. Moonshot argues no open-weights model should push the absolute frontier of offensive cyber capability, while wanting to raise what open models can do because they also serve defensive work like auditing legacy systems. The cited precedent is Anthropic's earlier finding that even GLM-5.3, which scores 45 on the index, can write exploits near Claude Mythos levels.

That trade-off is worth dwelling on. Deliberately degrading a capability before releasing weights is a policy choice with a measurable cost, and the 9.5-point drop on xdev-bench-hard is the visible price. A likely concern for adopters is that the cleaning is described at the data level rather than as a runtime filter, so the capability ceiling applies to everything built on the weights, including defensive security teams who would like the stronger version. Moonshot says it wants to raise that ceiling gradually, which implies future releases may move the line.

Why it matters: developers now have a model that is competitive with the top proprietary options on agentic and coding benchmarks, priced 33 percent lower per token, with weights promised by November 2026. Product teams building agents, CAD or game-engine tooling have a credible alternative to route work to, though per-task cost, internal benchmark caveats and the delayed open-weights date are reasons to test rather than assume parity. If the weights do ship on schedule, the open-model gap on this index would sit at 3 points, my inference being that procurement pressure on proprietary vendors intensifies well before that date.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Moonshot AI's Kimi K3.1 Pro scores 55 points on the Artificial Analysis Intelligence Index, an 11-point jump over its predecessor that puts it in fourth place. Only three Anthropic and Meta models rank higher. If Moonshot releases the weights as planned, the gap between open and proprietary models shrinks to just three points. The article Chinese AI model Kimi K3.1 Pro is closing in on Anthropic's Claude Opus 5.5 appeared first on The Decoder .