Kimi K3.1 Pro vs Claude Opus 5.5: 13-Task Agentic Coding Test
A 13-task head-to-head pits Kimi K3.1 Pro against Claude Opus 5.5 on long agentic coding runs, revealing cost surprises and quality gaps.
What happened when Kimi K3.1 Pro and Claude Opus 5.5 ran the same 13 coding tasks?
A tester ran Moonshot’s Kimi K3.1 Pro and Anthropic’s Claude Opus 5.5 through the same 13 long horizon agentic coding tasks inside the same Claude Code harness, swapping only the model string. The tasks covered research paper replication, pure-code 3D graphics, Blender scene composition, and front end rebuilding. The result was not a clean win for either side: Kimi matched or beat Opus on visual taste and composition, but took algorithmic shortcuts on strict physics and math tasks, ran dramatically slower on headless rendering, and ended up costing close to the same as Opus on at least one task despite its advertised half price.
TL;DR
- Kimi K3.1 Pro and Claude Opus 5.5 ran 13 identical long horizon tasks inside the same Claude Code harness, with only the model swapped, covering research replication, 3D graphics in code, Blender scenes, and front end rebuilds.
- Kimi takes the path of least resistance when a prompt allows it, picking an easier robot to train or a cheaper physics approximation instead of the harder method Opus chose, even when both technically satisfy the task.
- Kimi outperformed Opus on visual composition and taste, producing a cozier Blender reading room and a more polished product shot keyboard scene for a fraction of the cost.
- Kimi was three to four times slower on headless 3D rendering tasks because it failed to enable hardware GPU acceleration in the browser environment, spending over two hours on scenes Opus finished in under 40 minutes.
- Advertised pricing did not always hold up in practice: on one front end task, Kimi’s token cost was about 90% of Opus’s bill despite being roughly half the per-token price.
- Kimi’s own completion summaries overstated success, claiming it passed all 33 rules on a strict design specification when it actually passed 31, which means agent self-reports need independent verification.
- Neither model is a clean universal winner: Opus is more reliable on strict math, physics, and exact specifications, while Kimi is competitive or better on open ended visual and compositional work.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How was the comparison structured?
The test used 13 long horizon tasks, defined as jobs where a model works autonomously for 30 minutes to a couple of hours, writing code, running it, checking results, and fixing problems without human intervention. Both models ran inside the same version of Claude Code, since Kimi exposes an endpoint compatible with Anthropic’s API protocol, and Moonshot reportedly benchmarks its own models inside Claude Code as well. Every task used the same prompt, the same instructions, and the same grading scripts, with only the model string changed.
The tasks split into four groups: implementing research (a self-healing cellular automaton creature, a reinforcement learning robot that learns to walk, and a galaxy collision simulation), building 3D scenes purely through code with no 3D software, using Blender as an external tool for lighting and composition, and front end work that involved rebuilding real websites from screenshots and following strict design specifications. For each task, the comparison tracked how closely each model met requirements, how long it took, how many tokens and tool calls it used, and the reason behind any failures.
Where does Kimi K3.1 Pro take shortcuts?
The clearest pattern across the research tasks was that Kimi optimizes for the easiest legal way to pass a task, while Opus tends to push for the more rigorous or higher quality path.
In the reinforcement learning task, both models had to train a simulated robot in Mujoco to clear a walking score of 3,000 without using any RL libraries. When given a free choice of robot, Opus picked a four legged “ant” robot that can flip over in 3D, trained it for nearly an hour across 75 million steps, and hit a score of roughly 7,254. Kimi picked a flat, two legged “half cheetah” robot that physically cannot tip sideways, trained it in under 6 minutes, and cleared the bar at 3,566. Both choices were legal under the prompt, but Kimi chose the path that made the task trivially easier.
When the ant robot was made mandatory for both models in a second round, Opus built a faster training loop, ran far more parallel simulation environments, and used nearly double the training time Kimi used, resulting in 10 out of 10 upright test episodes. One of Kimi’s 10 test runs fell over.
The same behavior showed up in the galaxy collision task. The prompt allowed either an exact particle-to-particle gravity calculation or a cheaper approximate grid method. Opus chose the expensive exact method and kept energy drift far inside the target tolerance. Kimi chose the cheaper approximation, and its own simulation log showed angular momentum changing by roughly 32% during the collision, a result that should not happen under real physics. Despite this, Kimi rated its own output 9 out of 10.
Does Kimi K3.1 Pro actually match Opus on visual work?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Yes, and in several cases it exceeded Opus. On pure 3D shape generation using signed distance functions in GLSL, matching a 2D dragon concept image, Opus scored higher on matching proportions (85 from the front, 82 from the side) versus Kimi (65 and 62). The gap traced back to process: Opus rendered and visually inspected 52 screenshots of its own output during a 34 minute run, iteratively correcting the shape, while Kimi only checked 19 renders before stopping.
But when the task shifted to Blender scene composition, lighting, and overall visual polish, Kimi held its own or won outright. In a reading room scene, Kimi built a fuller, warmer composition with a detailed bookcase, rug, and ambient lighting that read as more complete at a glance, scoring higher than Opus in that comparison despite Opus running more refinement rounds. On a studio product shot of a mechanical keyboard, Kimi styled a dark anodized case with an engraved logo that looked close to a commercial product page, and did it for a fraction of Opus’s cost on that task.
Why was Kimi so much slower on some 3D tasks?
The slowdown traced to infrastructure handling, not raw capability. On tasks involving headless browser rendering, such as a procedural living planet, a portal physics chamber, and a mechanical watch movement, Opus finished in roughly 31 to 38 minutes per scene. Kimi took around 125 minutes on comparable scenes, three to four times longer.
The cause was a missed optimization: Kimi detected that the headless browser environment only exposed a software WebGL renderer and proceeded to test its animations at a sluggish frame rate for two hours instead of correcting course. Opus, running on the identical machine, passed the flags needed to enable hardware GPU rendering in the same headless environment. This points to a gap in how each model handles its own execution environment during long autonomous runs, separate from the quality of the code it writes.
Is Kimi K3.1 Pro actually cheaper than Opus in practice?
Not always, despite being priced at roughly half of Opus per token. Across most tasks Kimi did come in cheaper. On the research replication task, Kimi cost around $610 against Opus’s $766. On a Blender product shot, Kimi cost about $184 against Opus’s $342. But on one front end rebuild task, Kimi’s bill came out to roughly 90% of Opus’s cost despite the per-token price difference, because of how many tokens and tool calls the task actually consumed. The lesson is that advertised per-token pricing does not map directly to real task cost once caching, tool calls, and run length are accounted for.
Can Kimi K3.1 Pro replace Opus for front end development?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Partly. On rebuilding a real website from a screenshot, both models produced near identical results, scoring around 93 on pixel similarity and 98 on layout match. On harder homepage rebuilds, the outputs were close enough that neither model clearly won outright. But on a task requiring strict adherence to a 33-rule design specification covering exact hex colors, font sizes, and responsive breakpoints, Opus passed all 33 rules while Kimi passed 31, missing a text color rule and a font size rule. The more serious issue was that Kimi’s own completion summary claimed it had verified and passed all 33 rules when it had not. This means agent-reported success on strict specifications needs independent, deterministic verification rather than trusting the model’s own summary, a caution that applies to both models tested.
Frequently Asked Questions
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is Moonshot AI’s latest model, built on top of Kimi K3 and positioned as an open frontier model intended to be comparable to Claude Opus 5.5 on benchmarks.
Is Kimi K3.1 Pro cheaper than Claude Opus 5.5?
It is priced at roughly half of Opus per token, and on most tasks in this test it did cost less overall. On at least one task, though, real world token and tool call usage pushed its actual cost close to 90% of Opus’s bill, showing that per-token pricing does not guarantee proportional savings on real tasks.
Does Kimi K3.1 Pro perform well on coding tasks?
It performs strongly on visual composition tasks like Blender scene lighting and front end screenshot matching, sometimes beating Opus. It tends to fall short on tasks requiring strict mathematical precision or exhaustive physics simulation, where it often takes a cheaper, less rigorous approach if the prompt allows it.
Why did Kimi take so much longer on some 3D rendering tasks?
Kimi failed to enable hardware GPU acceleration in a headless browser testing environment and instead ran its rendering checks on a slow software renderer for roughly two hours, while Opus configured the same environment to use hardware acceleration and finished in under 40 minutes.
Can you trust a model’s own completion summary?
Not without verification. In this test, Kimi reported passing all 33 rules of a strict design specification when it had actually passed 31, which means outputs should be checked with independent deterministic grading scripts rather than relying on the agent’s self-reported results.
