Kimi K3.1 Pro vs GPT-6 Astra, Claude Opus 5.5 and DeepSeek: two tests worth a closer look
Adapted from WaWaGitHub's review, via SiAPI.ai
TL;DRKimi K3.1 Pro stood out for its SVG design and detailed interactions in a four-model review. Claude Opus 5.5 earned the reviewer's strongest praise for visual quality overall. DeepSeek V4.1 Flash 1017 had the lowest reported bill. The two cases below show why those judgments need some context.
WaWaGitHub compared the four models across five visual and 3D tasks, including a combined drawing challenge, a mechanical unicorn and a Blender stadium scene. This shorter account covers the first two cases and the overall spend. The observations come from that review; we have not repeated the tests.
The cases ask for two different kinds of success. An SVG can look convincing and still contain a wrong answer. A 3D page can have extensive controls while the object itself looks unfinished. Reading those results separately gives a more useful picture of Kimi than treating every task as a single contest with one winner.
Who got both the drawing and the math right?
Kimi K3.1 Pro and GPT-6 Astra were the only models to solve the embedded math problem correctly. The reviewer had combined three familiar drawing tests into one SVG brief, then added a candy-selection problem. The correct answer was 21.
Claude Opus 5.5 finished in three minutes, compared with 11 for GPT-6 Astra, 15 for Kimi and 17 for DeepSeek. But Claude and DeepSeek both missed the math. For a task that asks for a picture and a correct answer, that is a substantial failure even if the picture looks finished.
The reviewer preferred Kimi's composition and color palette. GPT-6 Astra also got the reasoning right, though its airplane had engines above the wing. Kimi came out of this case with the more convincing combination of visual judgment and a correct answer, at the cost of a longer wait than GPT.
Could Kimi match Claude's mechanical unicorn?
Claude Opus 5.5 produced the reviewer's favorite unicorn. The brief called for a procedural Three.js model with moving parts, two themes and an exploded view. It also ruled out imported 3D models and texture images, so each model had to construct the object in code.
Claude's result had a better-proportioned head, neck and hooves, with visible gears and metal trim that made the mechanism more believable. The reviewer also reported working gear ratios and inverse kinematics, the calculations that control how the legs move.
Kimi put more of its effort into the controls. Its model exposed 31 components, with hover readouts, values that could be pinned by clicking and a component selector. That made individual parts easier to inspect, although the reviewer felt the artwork still needed refinement.
GPT-6 Astra delivered working themes and basic interactions, but some gears sat outside the body. DeepSeek kept a consistent page style, yet rendered the whole unicorn black. The four runs took between 48 and 56 minutes. This case therefore offered little reason to choose on elapsed time alone; the differences were in the object and how people could interact with it.
How much did the full comparison cost?
The reviewer reported the following totals across all five tasks, including the three cases omitted from this excerpt.
| Model | Total spend |
|---|---|
| Kimi K3.1 Pro | $31.08 |
| GPT-6 Astra | $137.40 |
| Claude Opus 5.5 | $89.09 |
| DeepSeek V4.1 Flash 1017 | $23.96 |
DeepSeek's total used peak rates. These figures are the bills for this particular set of runs, rather than per-token prices or a forecast for another project. The review does not give enough billing detail to turn them into a controlled cost comparison.
Kimi's bill was much lower than Claude's or GPT's, and it had clear strengths in the two examples above. Whether that tradeoff suits a project depends on how much work remains after generation. A usable set of controls can save time; a model that needs substantial visual revision can give some of that time back.
Sources
WaWaGitHub's four-model review, translated on SiAPI.ai. Source for the reported outputs, elapsed times and total spend. Visual preferences are the reviewer's judgments.
For setup details, see our Kimi Coding Plan guide.