OSWorld
Integrity rank #42 of 61 · 8 models scored · top score 72.1 · Claude Sonnet 4.6
unknown
Strongest on discrimination, weakest on contamination resistance. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 72.1 | unverified· optimizedT1 | 2026-02-17 |
| 2 | Claude Opus 4.5 | 66.3 | unverified· optimizedT1 | 2025-11-24 |
| 3 | Kimi K2.5 | 63.3 | unverified· optimizedT1 | 2026-02-02 |
| 4 | Claude Sonnet 4.5 | 62.9 | unverified· optimizedT1 | 2025-09-29 |
| 5 | Claude Sonnet 4 | 43.9 | unverified· optimizedT1 | 2025-05-22 |
| 6 | Claude 3.7 Sonnet | 35.8 | unverified· optimizedT1 | 2025-02-24 |
| 7 | o3 | 23 | unverified· optimizedT1 | 2024-12-20 |
| 8 | Qwen2.5-72B | 5 | unverified· optimizedT1 | 2024-09-19 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
Wide gaps but few entrants — 67.1 points across just 8 models — so the benchmark discriminates, while the sample stays too small to call the ordering settled. This wide spread supports real separation between models, so rank differences are meaningful.
3 cited facts
An 80 with a B grade, ranked 34th of 51, is a middling health reading: the benchmark discriminates, but with enough friction that small cross-model gaps on it are better read as ties. Contamination is the weakest integrity component, narrowing how much to trust top scores. This score reflects the benchmark's health as a discriminator of frontier models under a disclosed harness.
6 cited facts
Plenty of test remains above the 72.1 leader — 27.9 points — so this benchmark can keep registering top-end progress, and an extended lead here reflects measurement rather than ceiling crowding. Only a single model sits at the top, so the ranking there still reflects genuine capability differences rather than noise.
5 cited facts
Scores from this harness are directly comparable due to consistent harness usage, meaning they can be used for head-to-head comparisons of task performance under the disclosed harness. The test set privacy status is unknown; therefore, it is unclear whether the test set is public or held out, and if it were public, high scores would deserve extra scrutiny for possible contamination. Contamination history is unknown, and the harness reports only model, version, performance, optimized flag, score source, benchmark release date, and model date via Epoch CSV, offering no qualitative assessment of contamination sensitivity.
4 cited facts
This benchmark does not document what it measures. Missing construction, assumptions, and limitations at the same time leaves a score with no way to interrogate it — the number can rank models, but it cannot explain anything. The sharpest caveat is the absence of documented limitations, as this prevents any assessment of the benchmark's potential to mislead.
5 cited facts
OSWorld: OSWorld as reported in Epoch AI's Capabilities Index CSV.
Claude Sonnet 4.6 leads OSWorld at 72.1. The full leaderboard above lists every recorded measurement, not just the headline number.
8 models have recorded scores on OSWorld, spanning a score spread of 67.1.
tensor.news grades OSWorld B for integrity (score 80/100), ranking #42 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — OSWorld still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on OSWorld are reasonably apples-to-apples.
Every OSWorld measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.