THE FORGE RANKINGS·2026-08-30

Claude Opus 4.8

Anthropic
6
RANK of 37 ranked
FERROX INDEX87.1 weighted mean of category positions
GRADEA front rank
EVIDENCE12 / 32 benchmarks measured
QUALITYTHIN Correlated evidence in: preference. Those category scores rest on benchmarks from a single family.

Not separable from GPT-5.4 (2026-03-05) at the measured 1.78 point band. The order is printed; the gap is not claimed.

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

30of 37 measured
30.4
13.1median 38.963.0

Coding

25of 37 measured
46.2
22.4median 50.070.7

Reasoning

13of 37 measured
77.5
18.1median 62.893.0

Preference

5of 15 measured
74.6
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.

APEX Agents

3/28
42.5% best 45.0% · Claude Fable 5
APEX CC-BY-4.0 max

FrontierMath

2/21
47.2% best 47.6% · GPT-5.4 (2026-03-…
Epoch AI CC-BY-4.0 max

LMArena Text (style-controlled)

6/20
1473 best 1507 · Claude Fable 5
LMArena CC-BY-4.0

FrontierCode

5/17
46.5% best 53.5% · Claude Fable 5
cognition.com CC-BY-4.0 claude-code max

LMArena WebDev

4/16
1539 best 1626 · Claude Fable 5
LMArena CC-BY-4.0

Terminal-Bench 4.0

5/10
23.6% best 51.8% · Claude Opus 5
Terminal-Bench Apache-2.0 claude-code max

Not measured

20/32
  • SWE-bench Verified (Epoch's own run)
  • Humanity's Last Exam
  • Terminal-Bench 2
  • 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
APEX Agents agents 42.5% 34.7 to 50.3 APEX CC-BY-4.0 / max source
LMArena Agent agents 0.025 16.9 to 33.2 LMArena CC-BY-4.0 source
Terminal-Bench 4.0 agents 23.6% 20.1 to 27.2 Terminal-Bench Apache-2.0 claude-code / max source
ALE-Bench coding 1412 not published ALE-Bench (Sakana AI with AtCoder) CC-BY-4.0 / none source
DeepSWE coding 51.8% 47.2 to 56.3 deepswe.datacurve.ai CC-BY-4.0 mini-swe-agent / high source
FrontierCode coding 46.5% not published cognition.com CC-BY-4.0 claude-code / max source
LMArena Text (style-controlled) preference 1473 81.2 to 82.5 LMArena CC-BY-4.0 source
LMArena WebDev preference 1539 66.5 to 68.3 LMArena CC-BY-4.0 source
ARC-AGI-2 reasoning 71.7% not published ARC Prize CC-BY-4.0 / medium source
FrontierMath reasoning 47.2% 41.5 to 53.0 Epoch AI CC-BY-4.0 / max source
GPQA Diamond (Epoch's own run) reasoning 88.4% 83.9 to 92.9 Epoch AI CC-BY-4.0 / low source
OTIS Mock AIME 2024-2025 reasoning 97.8% 93.4 to 100.0 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: claude-code, 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.