OSWorld 2.0
Integrity rank #27 of 65 · 9 models scored · top score 31.43 · Claude Opus 5
unknown
Strongest on harness comparability, weakest on discrimination. 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 Opus 5 | 31.43 | unverifiedT1 | 2026-07-24 |
| 2 | GPT-5.6 Sol | 27.34 | unverifiedT1 | 2026-07-09 |
| 3 | Claude Opus 4.8 | 20.6 | unverifiedT1 | 2026-05-28 |
| 4 | Claude Opus 4.7 | 18.2 | unverifiedT1 | 2026-04-16 |
| 5 | GPT-5.5 | 13 | unverifiedT1 | 2026-04-23 |
| 6 | Claude Sonnet 4.6 | 9.3 | unverifiedT1 | 2026-02-17 |
| 7 | Kimi K2.6 | 4.6 | unverifiedT1 | 2026-04-20 |
| 8 | MiniMax-M3 | 4.6 | unverifiedT1 | 2026-06-01 |
| 9 | Qwen3.7-Plus | 2.8 | unverifiedT1 | 2026-06-02 |
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.
This benchmark evaluates 9 models, with a best-to-worst spread of 28.63 points. That wide spread supports real separation between models, so rank differences are more likely to reflect genuine capability gaps rather than minor score noise.
3 cited facts
This benchmark's Benchmark Integrity Index is 87, graded A, and ranks 27th of 62 benchmarks. The weakest component of its integrity breakdown is discrimination, which means the benchmark separates frontier models less cleanly under its disclosed harness. The score therefore describes benchmark health as a discriminator of frontier models under a disclosed harness, not any model's capability.
6 cited facts
This benchmark is not saturated, with a top score of 31.43 and 68.57 points of headroom left to the ceiling. Only one model clusters near the top, so the field is not bunched within a couple of points. For the reader, that means there is still genuine room to separate the field at the top, so small ranking differences can reflect real capability gaps rather than noise.
5 cited facts
Scores for this benchmark are reported under a consistent harness, so they are directly comparable across the models covered, and the supporting notes document the model, version, performance figure, optimization flag, score source, release and model dates that frame each measurement. Whether the test set is public or held out is unknown, so the usual practical rule — that a public test set warrants more scrutiny for possible contamination than a held-out one — cannot be applied with confidence here. Because contamination history is unknown and the harness documentation does not characterize harness sensitivity, these figures should be read as task performance under a disclosed harness rather than as deployed capability, and any cross-harness comparison should be treated as provisional.
5 cited facts
What this benchmark is intended to quantify is not documented in the available fields, so its measurement target cannot be described here. How the test was assembled—its tasks, environments, and scoring procedure—is likewise not documented in the record provided. The assumptions underpinning its design and the limitations acknowledged by its authors are both unrecorded, as is any human baseline against which results could be read. The sharpest caveat for a reader is therefore that without documented measurement targets or acknowledged limitations, this benchmark's results cannot be checked for validity or meaningfully contrasted with an earlier benchmark or a related suite.
7 cited facts
OSWorld 2.0: OSWorld 2.0 as reported in Epoch AI's Capabilities Index CSV.
Claude Opus 5 leads OSWorld 2.0 at 31.43. The full leaderboard above lists every recorded measurement, not just the headline number.
9 models have recorded scores on OSWorld 2.0, spanning a score spread of 28.63.
tensor.news grades OSWorld 2.0 A for integrity (score 87/100), ranking #27 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — OSWorld 2.0 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 2.0 are reasonably apples-to-apples.
Every OSWorld 2.0 measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.