THE FORGE RANKINGS·2026-08-30

GPT-5.6 Luna

OpenAI
16
RANK of 37 ranked
FERROX INDEX70.8 weighted mean of category positions
GRADEB above the field median
EVIDENCE8 / 32 benchmarks measured
QUALITYFULL every category rests on independent benchmarks

Not separable from Claude Opus 4.6 and Gemini 3.1 Pro Preview 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

35of 37 measured
22.7
13.1median 38.963.0

Coding

28of 37 measured
43.9
22.4median 50.070.7

Reasoning

22of 37 measured
59.4
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

5/35
1667 best 2177 · GPT-5.6 Sol
ALE-Bench (Sakana AI with AtCoder) CC-BY-4.0 max

LMArena Agent

13/21
0.034 best 0.127 · Claude Opus 5
LMArena CC-BY-4.0

FrontierCode

11/17
39.8% best 53.5% · Claude Fable 5
cognition.com CC-BY-4.0 codex max

Terminal-Bench 4.0

8/10
17.3% best 51.8% · Claude Opus 5
Terminal-Bench Apache-2.0 codex max

Not measured

24/32
  • APEX Agents
  • FrontierMath
  • SWE-bench Verified (Epoch's own run)
  • LMArena Text (style-controlled)
  • Humanity's Last Exam
  • Terminal-Bench 2
  • LMArena WebDev
  • METR time horizons
  • 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
LMArena Agent agents 0.034 23.4 to 32.7 LMArena CC-BY-4.0 source
Terminal-Bench 4.0 agents 17.3% 14.4 to 20.1 Terminal-Bench Apache-2.0 codex / max source
ALE-Bench coding 1667 not published ALE-Bench (Sakana AI with AtCoder) CC-BY-4.0 / max source
DeepSWE coding 44.2% 41.3 to 47.2 deepswe.datacurve.ai CC-BY-4.0 mini-swe-agent / high source
FrontierCode coding 39.8% not published cognition.com CC-BY-4.0 codex / max source
ARC-AGI-2 reasoning 29.3% not published ARC Prize CC-BY-4.0 / high source
GPQA Diamond (Epoch's own run) reasoning 82.3% 77.0 to 87.7 Epoch AI CC-BY-4.0 / low source
OTIS Mock AIME 2024-2025 reasoning 66.7% 52.7 to 80.6 Epoch AI CC-BY-4.0 / low source

Configurations rolled up: 12. Rule: median observed configuration per benchmark (lower median, always a real measurement). Harnesses seen: codex, 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.