Kimi K3.1 Pro tested: 3D visuals, terminal coding and one fifth the cost

What did the review test?

The review, published on 302.AI's WeChat channel, ran K3.1 Pro through four batteries: logic and math (10 prompts), human intuition (7), coding simulation (12) and multimodal tasks (20). Each round was scored out of 10 and averaged across rounds. Coding ran in the 302.AI Studio client's Vibe mode, which drives a Claude Code sandbox, and every result is a first generation from one shared prompt set. The lab grades results from S down to C. This article summarizes their published findings. We have not rerun the tests.

Where does K3.1 Pro rank after the update?

Moonshot released K3.1 Pro on October 24. Per the review, the model is the K3 base retrained with 99 days of reinforcement learning across more than a hundred domains, followed by a one-third API price cut. On Artificial Analysis blind-test data quoted in the review, its AA Intelligence Index score is 55, up 11 points from K3, fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). The gap to the strongest closed model has narrowed from 12 points to 3. Moonshot plans to release the full weights by November 4, which the review expects to make it the strongest open-weights model. At review time it sat fifth on the Artificial Analysis overall board.

How strong are the 3D and visual results?

On BenchCAD, which tests rebuilding CAD geometry from multiple views, K3.1 Pro scores 92.7% with tools, close to Opus 5.5's 96.2% and GPT-6 Astra's 95.9%. Without any tools it scores 72.8%, nearly level with Opus 5.5's 73.0% and well ahead of Fable 5.1's 60.6%. The model writes and runs Blender and CAD scripts without MCP connectors. Kimi's site and app also ship a 3D Make feature that turns one sentence into a 3D model you can preview, edit and export for 3D printing.

Two interactive cases anchor the visual verdict. In a clay-style Halloween town, the reviewers praised the soft clay shapes, the three-layer sky gradient and a seeded, reproducible scene layout, and they flagged the mesh counts: each path segment spawns dozens of pebble meshes, enough to drop frames on slower devices. In a playable Quidditch game, K3.1 Pro shipped complete rules, distinct role behavior for keepers, beaters and seekers, and smooth flight controls, but a stadium of 4200 spectators with shadows and bloom strains low-end hardware, and there is no touch scheme for phones. GPT-6 Astra's counterparts showed cleaner onboarding cues, event sounds and far more readable code, yet stiffer geometry, instant day-night switching and thinner detail. On the SVG animation tasks K3.1 Pro built the more complex figures, and both models named all nine film directors in the multimodal check.

Can you trust the output?

On Kimi Hallu Bench Hard, the factual hallucination rate fell from K3's 23.4% to 8.3%. That beats Claude Opus 5.5's 15.0% in this comparison and trails only GPT-6 Astra's 5.5%. Office work points the same way: 64.4 on OfficeQA Pro v2, described as a SOTA result, and 37.0 on Moonshot's internal long-horizon business simulation, about 4.2 times K3's 8.9. Financial research rose from 58.8 to 70.4 and financial modeling from 50.4 to 57.6.

What about coding and cost?

On Terminal-Bench 4.0, which runs agents through terminal tasks under resource limits, the review reports 65.7% for K3.1 Pro against GPT-6 Astra's 58.2% and within a point of Claude Opus 5.5's 66.4%. K3 scored 13% on the same benchmark before the update.

The review also lists cache reads at $0.24 per million tokens for K3.1 Pro and cache writes at $12.50 per million for GPT-6 Astra. These are platform listings at review time, not our rate card; check the current pricing page before budgeting.

By those numbers the K3.1 Pro bill lands near one fifth of Astra's. Moonshot also sells Highspeed, which keeps the same capability with an advertised 2 to 3 times faster output generation.

To try Kimi K3.1 Pro, explore our Coding Plan.

Where does it still fall short?

The review is specific. The 3D output is expensive to render, with per-tile meshes pushing frame rates down on mid-range machines. The generated code leans on single-letter variables and CSS compressed into single lines, which hurts maintenance and second-stage development. The Quidditch build has no touch controls, and its strict catch detection steepens the learning curve. Across cases the reviewers write that dynamic implementation quality and delivery completeness still need work.

Should you try it for your next build?

For visual front-end work, 3D scenes and atmosphere-heavy demos, the review backs K3.1 Pro as the strongest pick in its comparison set, at a much lower bill. For long autonomous coding runs the benchmark gap to Opus 5.5 has nearly closed, but the caveats about polish and overhead are reasons to test rather than assume. Pick one of your own tasks with a result you can inspect, run it on K3.1 Pro and on your current model with matching settings, and compare quality, elapsed time and cost. Our earlier launch article covers the official rates and setup.