Eight independent tests from June to October 2026, covering Kimi K3, Kimi K3.1 Pro and Kimi Work.
01Strong
Three real coding jobs
Kimi K3 vs Claude Fable 5 · The New Stack · 20 Jul 2026
A bug fix, a refactor and a new feature in the open-source fd search tool, using the same prompts for both models.
- Both models produced fixes that passed every test (70/70 and 264/264).
- Kimi K3 cost $2.13 for all three jobs, against $5.98 for Fable 5.
- Kimi K3 took about 28 minutes in total, against about 7 for Fable 5.
Result: Same quality at about a third of the cost, but four times slower.
Read the source ↗
02Mixed
GitHub issues and trap tasks
Kimi K3 vs Claude Opus 4.8 · MindStudio write-up · 25 Jul 2026
Real GitHub issues from an active codebase, each planned, built and checked, then scored out of 70. A set of trap tasks targeted known weak spots.
- Simple issues: Kimi K3 scored within about 0.2 points of Opus 4.8's 64.3.
- Bigger builds: Opus averaged 62.2 and Kimi K3 stayed above 60.
- Trap tasks: Kimi K3 failed 36% of the time, against 8% for Opus.
Result: Close on clear work, much weaker on tricky, hidden-rule tasks.
Read the source ↗
03Good
Three app builds
Kimi K3.1 Pro vs Claude Opus 5.5 · AI Fire · 28 Oct 2026
The same prompts for a website copy, a task-management dashboard and a 3D racing game.
- Website copy: Kimi K3.1 Pro won, with a more accurate replica in 12 minutes against 20.
- Dashboard: Claude won, with clearer task details.
- 3D racing game: Claude won, with better graphics and track design.
Result: Kimi won 1 of 3. Best at copying designs, behind on complex builds.
Read the source ↗
04Good
Visual coding: K3.1 Pro vs K3
Kimi K3.1 Pro vs Kimi K3 · Overchat · 24 Oct 2026
Two browser apps: an aquarium simulation and a low-poly 3D island with a day-night cycle.
- K3.1 Pro wrote richer code (about 33,000 characters against 19,000) and a more detailed daytime island.
- But it missed one requirement: the night sky had no stars, while K3 included them.
- K3.1 Pro took 10.6 and 11.6 minutes, against 3 and 4.1 for K3.
Result: More detail and accuracy overall, but two to three times slower, and still worth checking.
Read the source ↗
05Good
A real 3D workflow
Kimi K3.1 Pro vs Claude Opus 5.5 · InfoQ developer test · Oct 2026
A developer used K3.1 Pro in a real 3D modelling workflow soon after launch.
- K3.1 Pro produced usable 3D work.
- It didn't finish in one prompt and needed several rounds of asking it to keep improving.
- Claude Opus 5.5 had produced a more complete result first time in the developer's earlier test.
Result: Capable at 3D, but expect to iterate.
Read the source ↗
06Good
Kimi Work desktop tasks
Kimi Work beta · Machine Heart · 4 Jun 2026
Five jobs at launch: an industry financing list in Excel, sorting local files, posting through a logged-in browser, 30 long speeches, and a stock analysis.
- Parallel agents built the financing list and an Excel summary.
- The 30-speech job produced a 104-page Word file in just over 50 minutes. It hit one timeout and restarted on its own.
- The stock analysis made 20 tool calls and wrote an 11-page report.
Result: Real end-to-end work on the desktop, with beta-stage hiccups.
Read the source ↗
07Good
Video understanding
Kimi K3 API · Kimi-AI.chat · 29–30 Jul 2026
300 fixed questions about 50 short test videos, sent through the Kimi API.
- Kimi K3 answered 222 of 300 questions correctly.
- That is 74.0% on this test set.
- The test didn't cover the Kimi app, long videos or audio.
Result: Solid video understanding, with one in four answers wrong.
Read the source ↗
08Strong
Independent benchmark index
Kimi K3.1 Pro · Artificial Analysis · Oct 2026
A composite index of reasoning, knowledge, maths and coding tests, run independently.
- K3.1 Pro scored 55, fourth overall, up 11 points from K3's 44.
- Only Claude Opus 5.5 (58), Claude Sonnet 5.5 and Muse Wonder Full ranked higher.
- It used 150M output tokens on the index, against an 87M median for its price tier, at $2.11 per task.
Result: Near the top on quality, but slower and wordier than most.
Read the source ↗