Kimi K3.1 Pro: Pat Simmons Tests Three Builds Against Frontier Models

Moonshot describes Kimi K3.1 Pro as a model for full-project coding and polished 3D work. Pat Simmons of AI for Mortals put that claim beside four rivals: Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra and DeepSeek V4.1 Flash. His video compares three demanding artifacts: a playable billiards game, a product-site rebuild and a recreation of Moonshot's launch film. The useful question is not whether one model wins a marketing headline. It is whether the delivered artifact works, resembles the brief and justifies its time and token use.

Watch the Three Tests

Jump directly to the billiards reveal, website reveal or launch-video reveal. The player above is Pat's review; the official Moonshot film appears alongside the video-rebuild task below.

Same Briefs, Different Agent Environments

Pat reports that each model received the same prompt per task, ran unattended, and was not steered during the build. That is a useful controlled brief, but the harnesses were different: Kimi ran in Kimi Code on a subscription; Opus and Fable in Claude Code; Astra in Codex; and DeepSeek through Claude Code with OpenRouter. Differences in browsing, tool calls, execution, caching and available integrations can affect both the result and the bill. This is a creator-run set of builds, not a standardized multi-seed model benchmark.

The companion article describes a fourth, parametric Sagrada Família test, but it is not among the three result segments in the supplied video outline. Its results were also still pending in the article. We have not folded it into this three-build comparison.

Test 1: Could It Build a Game, Not Just a Screenshot?

Pat's original brief calls for a browser-based first-person pool game named CLUB No.9 using Three.js and Rapier. The model must make a convincing late-night room and a full 8-ball rules loop, not merely render a table. Players should aim and charge shots with the mouse, strike different points on the cue ball to change spin, walk around for a better view, practice freely, face a computer opponent and replay shots. Balls need plausible sliding, rolling, cushion rebounds and pockets that only accept balls that genuinely enter. Sound, a controls screen, a runnable project and README are also part of the deliverable.

That prompt is intentionally demanding. A pretty frame can hide broken interaction, while working controls can hide toy physics. To reproduce the evaluation, inspect both the visual output and a full playthrough: take several shots with top, bottom and side spin; test fouls and ball-in-hand; complete a game; replay a shot. Pat makes the five builds available to open in a browser, so readers can inspect more than the edited reveal. His article contains the exact copyable prompt; this is a condensed description of it.

What the Published Billiards Ledger Actually Says

Figures are Pat's published billiards measurements, not our reruns. He explicitly describes the dollar amounts as API-equivalent: real token counts priced at published API rates, even where the session itself ran under a subscription. Kimi's $3.46 is therefore not a separate subscription invoice. DeepSeek was cheapest in this ledger; Kimi was fastest. Neither column alone identifies the best finished game. The blind ranking and reveal supply the visual and interaction judgment.

Also notice how widely token use varies even with the same brief. That is part of the cost story: the tool loop, model verbosity and cache behavior can matter as much as the advertised input/output price per million tokens. Before adopting a cheaper model for a production workflow, compare the cost of an accepted artifact, including inspection, corrections and reruns.

Test 2: Rebuild a Site Whose Identity Is Motion

The second brief asks for a close rebuild of MORPHO's product site. The visual challenge is the metal can, landscapes and scroll-driven transitions; the functional challenge includes the intro, section navigation, Blue Zones selector, hovers and mobile layouts. Pat's Clone App Pat Pro repository makes the method explicit: inspect live states, extract computed styles instead of guessing, build from a design system, run style assertions and revisit visual differences. For this test Pat instructed agents to create their own imagery, 3D and footage rather than reuse the source site's assets.

Adapted evaluation brief: map the site at desktop and mobile widths, record type/spacing/color/motion values, reproduce the can and interactive sections with independently created assets, then run the repository's assertion gate and compare identical viewports. A screenshot-only clone would fail the core of this brief. See the complete original prompt for the exact constraints and finish criteria.

The video walks through the builds from 19:44 and the ranking at 29:11. The companion article still shows pending for each website build's token count, cost and runtime. We therefore do not manufacture a website price table or treat a visual ranking as an audited timing result.

Test 3: Recreate the Kimi Launch Film

Here the reference is Moonshot's approximately 94-second, 1920 × 1080, 25 fps launch video. Pat asked each agent to break it into shots, match the edit and visual language, render a new film, and check it side by side. Crucially, the task forbids reusing reference frames, footage or audio in the final render. The Launch Video Clone repository supplies the workflow; the full source prompt specifies timing, 3D depth, typography, camera motion, sound and final deliverables.

Watch the original film below, then jump to Pat's rebuild comparison and ranking. A fair review checks whether shots match in order and duration, whether the generated imagery reads as a coherent scene rather than a static card, and whether the final MP4 really plays with sound. It should also credit time spent rendering and correcting: the article's launch-video cost and duration cells were still pending when checked.

Pricing, Benchmarks and What “Open Weight” Means Today

Moonshot says K3.1 Pro has a 1-million-token context window and is available through Kimi.ai, Kimi Work, Kimi Code and its API. The separate Highspeed option is advertised at two to three times the speed. Moonshot calls K3.1 Pro open weight, but says the full weights will arrive by November 4. As of this post's date, the hosted access exists; a local-weights download is still a promised release. Our setup guide separates browser, desktop, coding and API paths.

Pat's article quotes $2 input / $10 output per million tokens for Kimi's API and compares that list price with the rivals. Check Kimi Open Platform before spending: cached reads, output volume, high-speed tiers and the chosen agent harness can change a real bill. Moonshot's own Terminal-Bench 4.0 result is 65.7%, compared with 66.4% for Opus 5.5 and 58.2% for Astra in its published chart. Those are vendor-reported benchmark figures, not Pat's three-build scores, and should not be used as a substitute for inspecting his actual games, sites and films.

A More Useful Verdict Than “Best Open Model”

This review is valuable because the assignments expose different failure modes. Billiards tests physics and complete interaction; the website tests visual fidelity plus live behavior; the film tests timing, scene craft and rendered delivery. Kimi's published billiards bill is strikingly low beside Opus's, but this is a single, mixed-harness comparison. The website and film tables are not yet a complete public cost ledger. Do not turn one quick build into a claim that Kimi is categorically faster, cheaper and better on every kind of work.

For a practical trial, borrow one of Pat's exact prompts, define acceptance tests before starting, run the same task through the tools you actually use, and record wall time, token classes, retries and human cleanup. Inspect the artifact yourself. That is the comparison worth paying attention to.

Video Chapters

Original Prompts, Builds and Sources

Primary review: Pat Simmons's full video and the AI for Mortals companion article, which contains the copyable prompts, billiards measurements and pending status for the other result tables. Artifacts and methods: the live billiards showcase, website-clone skill and launch-video skill. Vendor context: Moonshot's K3.1 Pro release and API platform. Rankings are Pat's judgments in his video; published token and cost numbers are his measurements; Moonshot benchmark claims are labeled separately. We have not rerun the five models.