THE FORGE RANKINGS·2026-08-30

Kimi K3

Moonshot
4
RANK of 37 ranked
FERROX INDEX93.7 weighted mean of category positions
GRADEA+ top of the measured field
EVIDENCE8 / 32 benchmarks measured
QUALITYFULL every category rests on independent benchmarks

Not separable from Claude Fable 5 at the measured 1.78 point band. The order is printed; the gap is not claimed.

No measurement in Preference. That weight was redistributed, which is the same as assuming this model would have scored its own average there. It is an assumption, not a measurement.

Compared to what

Each rail is one classification. Every ranked model measured on it is a tick; this model is the orange marker. A hatched rail means this model has no measurement in that classification and its weight was redistributed.

Agents

14of 37 measured
43.4
13.1median 38.963.0

Coding

17of 37 measured
52.1
22.4median 50.070.7

Reasoning

11of 37 measured
80.1
18.1median 62.893.0

Preference

no qualifying measurement
53.1median 71.882.5

Every measurement, and the field behind it

One card per benchmark this model has been measured on. The rail shows where it sits against every other ranked model measured on the same benchmark. Where the evaluator keys on a harness, the harness is named, because the same model scores differently under a different one.

ALE-Bench

7/35
1524 best 2177 · GPT-5.6 Sol
ALE-Bench (Sakana AI with AtCoder) CC-BY-4.0 max

APEX Agents

5/28
39.3% best 45.0% · Claude Fable 5
APEX CC-BY-4.0

LMArena Agent

4/21
0.092 best 0.127 · Claude Opus 5
LMArena CC-BY-4.0

DeepSWE

5/17
68.5% best 72.8% · Claude Opus 5
deepswe.datacurve.ai CC-BY-4.0 mini-swe-agent max

FrontierCode

6/17
44.2% best 53.5% · Claude Fable 5
cognition.com CC-BY-4.0 mini-swe-agent none

Not measured

24/32
  • FrontierMath
  • SWE-bench Verified (Epoch's own run)
  • LMArena Text (style-controlled)
  • Humanity's Last Exam
  • Terminal-Bench 2
  • LMArena WebDev
  • METR time horizons
  • Terminal-Bench 4.0
  • Aider Polyglot
  • OSWorld
  • Cybench
  • BrowseComp
  • IMOAnswerBench
  • SWE-bench Multilingual
  • AIME
  • CyberGym
  • HMMT
  • MCP-Atlas
  • SWE-bench Pro
  • Tau2-Bench Airline
  • Tau2-Bench Banking
  • Tau2-Bench Retail
  • Tau2-Bench Telecom
  • Tool-Decathlon
A missing benchmark is not a zero and is never scored as one.

Check it yourself

Every number on this page, with who measured it, the interval they published, the harness it was run under, and where to go and read it. Nothing here was measured by Ferrox Labs.

BenchmarkClassPublishedInterval (normalised) Measured byLicenceHarness / effort
APEX Agents agents 39.3% 31.5 to 47.1 APEX CC-BY-4.0 source
LMArena Agent agents 0.092 45.4 to 49.5 LMArena CC-BY-4.0 source
ALE-Bench coding 1524 not published ALE-Bench (Sakana AI with AtCoder) CC-BY-4.0 / max source
DeepSWE coding 68.5% 64.0 to 73.0 deepswe.datacurve.ai CC-BY-4.0 mini-swe-agent / max source
FrontierCode coding 44.2% not published cognition.com CC-BY-4.0 mini-swe-agent / none source
ARC-AGI-2 reasoning 55.0% not published ARC Prize CC-BY-4.0 / high source
GPQA Diamond (Epoch's own run) reasoning 91.9% 88.1 to 95.7 Epoch AI CC-BY-4.0 / high source
OTIS Mock AIME 2024-2025 reasoning 93.3% 86.0 to 100.0 Epoch AI CC-BY-4.0 / high source

Configurations rolled up: 6. Rule: median observed configuration per benchmark (lower median, always a real measurement). Harnesses seen: mini-swe-agent.

Snapshot 2026-08-30-62e4e85f043b, manifest dc46131315046f1e. Grades are positional: position in the measured field, as a percentile of rank among ranked entries, n=37.