World of AI's Kimi K3.1 Pro test moves quickly from benchmark slides to a browser editor, two games, animated 3D scenes, a long research report and computer use. Its strongest case is visual prototyping. Its most revealing moment is less flattering: the zombies move, but the shooting control does not work in the presenter's run.
K3.1 Pro looks compelling for creative coding and 3D-first experiments, but the video does not establish that it replaces Opus 5.5 across engineering work. Standard first-party API list rates are half Opus 5.5's, not universally five times cheaper. A quoted faster pagoda build is one example, not a general throughput test. Kimi serves the model today; full weights are promised by November 4, 2026.
Watch the Full Test
How to read this review: The observed builds and quality judgments belong to World of AI; we did not rerun them. Kimi's release and benchmark claims are vendor data. The video also contains a Kimi promotional offer; the supplied destination is truncated, so we link to the official API instead of guessing the campaign URL.
What the Demos Actually Show
On screen
A Canva-style vector editor and an animated landing page show usable structure, typography and motion. The presenter particularly likes the design cohesion.
Limit
These are selected demos. The video does not include repeat runs, accessibility checks, responsive screenshots or maintained-code review.
On screen
A samurai-themed browser scene has character movement and textured scenery; a zombies prototype has enemy animations and weapon motion.
Limit
During the zombies demonstration, the presenter cannot fire even after buying another gun. Attractive output is not a complete, tested game. The builds are inspired by well-known games, not licensed reproductions of their assets.
On screen
The presenter prefers Kimi's running-horse animation over the displayed Opus and Astra outputs. He also compares a Three.js skyline, an Arc de Triomphe scene, a foldable-phone render and an animated SVG peacock.
Limit
Visual preference is subjective, and the video does not provide a reproducible prompt, harness and full run logs for each comparison. One phone scene is reported at about 90 minutes and $12.50, not instant or free.
On screen
A 114-page quantum-computing report contains charts and citations; a computer-use demo draws a portrait in Paint. The video also discusses a weak robotics result.
Limit
Length and cited links do not prove a report's facts or financial figures. The shown robotics score is a benchmark result, not proof of reliable physical-world execution. The painting is a single visual task.
For a separate hands-on run that catches more failure modes, read our Matthew Berman K3.1 Pro review. His Rubik's Cube works well, while his accuracy-sensitive phone, climbing knot and benchmark graphic do not.
The Benchmark Picture Is Strong, Not Absolute
Kimi's release reports 65.7% on Terminal-Bench 4.0 for K3.1 Pro, versus 66.4% for Opus 5.5 in its comparison. Its with-tools BenchCAD subset reports 92.7% for Kimi versus 96.2% for Opus. These are vendor-published comparisons with benchmark-specific harnesses and reasoning settings; “beats Opus” cannot be generalized from the count of winning rows in a launch chart.
Artificial Analysis independently places K3.1 Pro at 55 on its Intelligence Index as of October 25, ranked fourth on its launch assessment. That is an index score, not 55% accuracy. Its launch write-up reports a different Terminal-Bench run at 67%; it should not be merged with Kimi's 65.7% as though both were the same test configuration. The presenter himself says Opus and Astra may still outperform Kimi on demanding agentic coding.
Cheaper, Faster, or Both?
| First-party API model | Input / 1M tokens | Output / 1M tokens | What changes |
|---|---|---|---|
| Kimi K3.1 Pro | $2 | $10 | Standard Kimi tier |
| Kimi K3.1 Pro Highspeed | $4 | $20 | Kimi claims 2-3x standard Kimi speed |
| Claude Opus 5.5 | $4 | $20 | Standard Anthropic tier |
Prices above are the Kimi API and Anthropic API list prices checked October 25. Standard Kimi is 50% less per input and output token than standard Opus; Kimi Highspeed lists the same input and output rates as standard Opus. Cache behavior, reasoning tokens, tool calls and retries can change the actual bill. Subscription plans are a different purchase.
For one animated voxel pagoda, the video reports Kimi at 15 minutes and $5.71 and Opus at 32 minutes and $17.13. That is about three times cheaper and roughly twice as fast for this demonstrated build. Without the full prompts, settings, token usage and acceptance criteria, it is not a universal head-to-head. Artificial Analysis measures Kimi's standard API at roughly 73.5 output tokens per second, below the median on its model page; speed depends on whether you mean output rate, first-token latency or completed-task time.
How to Try It, and What Is Not Yet Available
K3.1 Pro is currently available through Kimi's app, Kimi Work, Kimi Code and the API. Kimi says the full weights will be released by November 4, 2026. As of this article, “open-weight” is its announced distribution plan, not a downloadable package to self-host. Kimi calls it 3T-class; the underlying K3 architecture has 2.8 trillion total parameters, not three trillion active per token. The one-million-token context window is a maximum, not a promise that million-token sessions are cheap or accurate throughout.
Our Take
Test K3.1 Pro first on an acceptance-tested 3D or front-end task where visual quality and API cost matter. Include a functional check, not just a screenshot, and compare total cost after fixes. Keep Opus in the comparison for complicated agentic engineering rather than assuming one creative win settles it. For long financial reports, inspect every material claim against the cited primary document before sharing the result.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | Introduction and open-weight release plan |
| 00:49 | Kimi-published benchmark comparisons |
| 01:38 | Finance and office-work scores |
| 02:09 | Artificial Analysis Intelligence Index |
| 02:33 | API pricing |
| 02:54 | Highspeed variant |
| 03:01 | How to access Kimi |
| 03:50 | First impressions |
| 04:27 | Front-end development |
| 04:43 | Canva-style editor |
| 05:01 | Animated landing page |
| 05:41 | Ghost of Tsushima-inspired browser build |
| 06:41 | Zombies game and broken shooting control |
| 08:32 | Where Kimi and Opus differ |
| 09:47 | Horse animation comparison |
| 10:39 | Three.js skyline |
| 11:29 | Blender, CAD and BenchCAD |
| 12:29 | Arc de Triomphe model comparison |
| 13:19 | 3D creator verdict |
| 14:04 | 114-page financial research report |
| 15:24 | Foldable-phone Blender and web demo |
| 16:22 | Animated SVG peacock |
| 17:02 | Voxel pagoda cost and time comparison |
| 17:45 | Robotics benchmark weakness |
| 18:33 | Computer-use painting demo |
| 19:00 | Final assessment |
Sources and Scope
Video evidence: World of AI's full review and the supplied transcript. Primary specifications and prices: Kimi's K3.1 Pro release, Kimi API and Anthropic's Opus 5.5 model page. Independent evaluation: Artificial Analysis model measurements and its launch analysis. We have not audited the report's 114 pages, reproduced the builds or verified the truncated X post URLs in the video description. Figures and access can change; checked 25 October 2026.