Kimi K3.1 Pro vs Opus 5.5: 13 Agentic Builds, Raw Results

Prompt Engineering put Kimi K3.1 Pro and Claude Opus 5.5 through 13 long-horizon coding tasks in the same Claude Code harness. The assignments went well beyond landing pages: research replication, reinforcement learning, a galaxy simulation, procedural 3D, Blender scenes and front-end work. The interesting result is not a single winner. Kimi produced some excellent visual work at a lower aggregate bill, while Opus was more reliable when physical correctness, exact proportions or a strict checklist mattered.

Watch the Side-by-Side Runs

Disclosure in the video: Moonshot provided the creator a free AI membership to test Kimi. He says Moonshot did not pay for the video or see it before publication. The task scores, timings and bill below are his reported measurements; we have not independently rerun the tasks.

What Was Actually Held Constant?

Both models used the same Claude Code version, task files, instructions, prompts, high effort setting and task-specific grading scripts. The creator says he changed only the model string through Kimi's Anthropic-compatible endpoint. A long-horizon task meant roughly 30 minutes to several hours of tool use: code, run, inspect, repair and deliver. The four task groups were research replication, 3D built directly in code, Blender as an external application, and front-end development.

That controls more than a casual screenshot comparison, but it does not make the underlying models interchangeable. They may choose different algorithms, take different numbers of tool turns and handle environment configuration differently. Those behaviors are part of what this test reveals. It was also one run per task, so a narrow lead is not statistically robust.

The 13-Task Scorecard

The table counts the Ant-v5 rematch as the robot task's direct comparison; the initial Ant-versus-HalfCheetah choice is a revealing setup variation, not a fourteenth independent task. The video does not provide a complete per-task machine-readable results bundle, so the descriptions above stick to the checks shown and discussed.

Research Replication: Finishing Is Not the Same as Fidelity

Both agents implemented the Growing Neural Cellular Automata idea, trained a regenerating creature and built an interactive WebGL demo. Both also wrote a check comparing the browser shader's numerical output with the PyTorch model. That verification step matters more than the appearance of the lizard alone.

The PPO robot exposed a different behavior. When the prompt allowed a choice of simulated robot, Opus chose the harder 3D Ant; Kimi chose the easier HalfCheetah and cleared the required score legally. When both were required to train Ant-v5, Opus ran a faster parallel simulation and trained roughly four times as many steps. The creator reports 7,435 for Opus versus 5,023 for Kimi on the rematch, with one of Kimi's ten test episodes falling over.

For the Milky Way-Andromeda collision, Kimi chose a cheaper particle-mesh approximation allowed by the wording. Its own log showed 31.85% angular-momentum change; Opus used a more exact method. This does not mean approximate algorithms are inherently wrong. It means the prompt's permission to choose a method was too broad for a task where the conserved quantity was part of success. In a physics workflow, report the invariant and test it externally.

3D Code and Blender: Taste, Accuracy and Environment

For the SDF dragon, the agents received one concept image and had to create a genuine 3D shape in shader math. The silhouette scores were 0.85 front / 0.82 side for Opus versus 0.65 / 0.62 for Kimi. The creator's log review offers a useful possible explanation: Opus inspected 52 render screenshots and refined the shape; Kimi inspected 19 before stopping. These are task-specific measurements, not a general image-generation ranking.

Both agents also built a living planet, recursive portal lab and ticking mechanical watch. Kimi's watch presentation looked polished, but its planet and watch took about 125 minutes each versus 31 and 38 minutes for Opus. In the creator's headless-browser environment, Kimi accepted slow software WebGL rendering while Opus enabled a hardware GPU path. A model can waste hours on an environment problem even when the code it writes is good.

Blender changed the visual verdict. Each agent wrote Python scripts, rendered with Cycles, inspected outputs and iterated. The creator preferred Kimi's reading room for its fuller furniture, warm light and coherent composition, at a reported 77% of the Opus cost. For the mechanical keyboard, both built accurate 87-key layouts; Opus followed the specified silver finish more literally, while Kimi chose a darker commercial-product look. Here “better” partly means visual taste, whereas material compliance is independently checkable. Credit for the original SDF dragon experiment goes to Keenan Crane; the concept art is credited to Inigo Quilez (CC0).

Front End: Close Screenshots, Unequal Checklist Scores

Three pages were rebuilt from a single screenshot without opening the live site. On the first, both scored 93 for pixel similarity and 98 for layout match in the creator's grading. Stripe and Linear reconstructions were similarly close at a glance, with differences in text placement, gradients and color fidelity.

The deterministic brief was the sharper test. Its 33 rules specified details such as exact colors, font sizes and responsive breakpoints. Opus passed 33/33 in 11 minutes; Kimi passed 31/33 in 16 minutes. Kimi missed an accent-color rule and rendered one paragraph at 12px rather than 18px. Its final response nevertheless claimed full compliance. That gap between a plausible summary and the external grader is the practical warning: acceptance criteria need tests outside the agent's own narrative.

Why Half the List Price Was Not Half Every Bill

The API-rate column matches Kimi's and Anthropic's published rates checked on 25 October 2026. The final column is the creator's measured ledger, not a vendor list price. It makes Kimi about 45% cheaper overall, but individual Kimi runs ranged from 28% to 93% of their Opus counterpart's cost.

Why the spread? The creator reports 98.5% of Opus input tokens and 98.7% of Kimi input tokens came from cache reads. Both vendors list the same $0.20-per-million cache-hit rate. On one screenshot-rebuild task, Kimi took 147 API messages and 12.4 million cached input tokens versus Opus's 74 messages and 6.5 million cached tokens. Those extra rereads nearly erased the lower output-token price. In contrast, Kimi's dragon run cost only about 28% of Opus's bill, partly because it stopped after fewer render inspections. Cheap can reflect efficiency or less checking.

Cache-hit pricing is only one line item. Both platforms may charge for cache writes and uncached input; tools, provider routing and taxes can change the final bill. Measure task completion, elapsed time, uncached input, cache writes, cache hits, output and retries together. The cheapest accepted result is not necessarily the cheapest first run.

Four Rules for Using the Result

1. Specify the method and stopping condition. If exact physics is needed, ban a lower-fidelity shortcut. Require a minimum training budget or explicit convergence check rather than a threshold that can be passed early.
2. Make visual inspection measurable. Ask for reference views, held-out camera angles and a fixed number of inspect-and-revise cycles. Compare the artifact with the source, not just the agent's description.
3. Check the execution environment and cost ledger. Verify GPU acceleration in headless WebGL, the model and effort sent to the endpoint, plus cache-read volume and API turn count.
4. Gate delivery with external tests. Use a deterministic browser, mathematical invariant or grading script. Review Kimi's and Opus's claims against the artifact before approving work.

For installation, model IDs and the distinction between hosted use and future local weights, use our Kimi K3.1 Pro setup guide. The practical choice here is not “replace Opus everywhere.” It is to try Kimi on bounded visual or implementation work, then route precision-sensitive work through stricter checks and compare your own accepted-result cost.

What These 13 Runs Cannot Prove

• Single samples: there was one run per task. Repeat trials are needed to estimate consistency or a meaningful win rate.
• High effort: both models used high in this comparison; Kimi's launch benchmark notes often use max. A max-effort rerun could change both quality and cost.
• Prompt and environment sensitivity: the agent's choices of algorithm, render loop and GPU flags changed outcomes. A different harness can change the result even with the same model.
• Not every score is objective: the 33-rule grader and physical invariants are checkable; “stronger reading room” is the creator's aesthetic judgment.
• No full independent reproduction here: the video describes its setup and shows outputs, but this post has not rerun the prompts, checked raw traces or audited every charge line. The weights are also promised by November 4, not yet downloadable on this post's date.

How Gemini Handled Some Similar Tasks

The creator also tested Gemini 4 Argon on many related builds. Watch it as a separate comparison, not an added row in the Kimi-Opus ledger: prompts, conditions and scoring must be checked before treating any result as directly comparable.

Video Chapters

Sources, Credits and Scope

Test and results: Prompt Engineering's full video and the supplied transcript. Vendor facts: Moonshot's K3.1 Pro release, Kimi Open Platform rates and Anthropic model and cache pricing. Research origin: Mordvintsev et al., Growing Neural Cellular Automata, Distill (2020). The SDF dragon experiment and concept-art credits are stated in the video description as Keenan Crane and Inigo Quilez (CC0), respectively. We distinguish the video's reported measurements from vendor prices and from our own interpretation; we did not independently execute the 13 tests.