OSWorld 2.0
Integrity rank #39 of 61 · 7 models scored · top score 20.6 · Claude Opus 4.8
unknown
Strongest on contamination resistance, 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 4.8 | 20.6 | unverifiedT1 | 2026-05-28 |
| 2 | Claude Opus 4.7 | 18.2 | unverifiedT1 | 2026-04-16 |
| 3 | GPT-5.5 | 13 | unverifiedT1 | 2026-04-23 |
| 4 | Claude Sonnet 4.6 | 9.3 | unverifiedT1 | 2026-02-17 |
| 5 | Kimi K2.6 | 4.6 | unverifiedT1 | 2026-04-20 |
| 6 | MiniMax-M3 | 4.6 | unverifiedT1 | 2026-06-01 |
| 7 | 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 scores seven models, with a spread of 17.8 points between the best and worst results. That wide spread supports real separation between models, so rank differences are likely to reflect genuine capability gaps.
2 cited facts
This benchmark's Benchmark Integrity Index is 82, a B grade, ranking 39th among 61 benchmarks in the universe. The weakest integrity component is discrimination, which limits how sharply the benchmark separates frontier-model scores. The score therefore communicates the benchmark's health as a discriminator under its disclosed harness, not any model's capability.
6 cited facts
This benchmark is not saturated, with a top score of 20.6 and 79.4 points of headroom remaining. Only one model clusters near the top, placing the leading band within a couple of points. Because the field is not saturated, small differences among the current leaders are not yet noise, leaving genuine room to separate the top competitors.
5 cited facts
Scores from this benchmark reflect task performance under a disclosed harness; because harness comparability is labeled "consistent harness", results are directly comparable across runs cited in the same source. Test-set privacy is unknown, so it is not possible to determine whether the test set is public or held out, meaning the appropriate level of scrutiny for possible contamination cannot be established. Contamination history is unknown as well, so no prior contamination incidents can be cited; in practice, any high score should still be interpreted as harness-specific task performance rather than deployed capability.
5 cited facts
This benchmark's intended measurement is not documented in the provided record, so its precise scope cannot be stated. How the test was built is likewise undocumented, leaving its construction process unspecified. The benchmark's underlying assumptions and limitations are also not documented, meaning any caveats about where it may mislead are unavailable from this record. The sharpest caveat a reader should hold is that because both assumptions and limitations are undocumented, this benchmark offers no documented basis for judging where its results do not apply.
6 cited facts
OSWorld 2.0: OSWorld 2.0 as reported in Epoch AI's Capabilities Index CSV.
Claude Opus 4.8 leads OSWorld 2.0 at 20.6. The full leaderboard above lists every recorded measurement, not just the headline number.
7 models have recorded scores on OSWorld 2.0, spanning a score spread of 17.8.
tensor.news grades OSWorld 2.0 B for integrity (score 82/100), ranking #39 of 61 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.