Frontier Log

Claude models on OSWorld-Verified

Changed by Anthropic on

Said it had changedPrinted a new figure

Anthropic re-ran older models on OSWorld-Verified after fixing a bug in its zoom tool and raising the limit on tokens per turn from 16K to 128K. Claude Opus 4.7 rose from 78.0% to 82.8%, Claude Opus 4.6 from 72.7% to 78.1%, Claude Sonnet 4.6 from 72.5% to 78.4%, and Claude Mythos Preview from 79.6% to 85.4. The Claude Sonnet 5 post later printed 78.5% for Sonnet 4.6. The Claude Opus 4.8 system card says: “we have been underreporting OSWorld performance across our model families”.

Anthropic Raised Its OSWorld-Verified Scores. OSWorld-Verified: Claude Opus 4.7 78.0% before and 82.8% after, Claude Opus 4.6 72.7% before and 78.1% after, Claude Sonnet 4.6 72.5% before and 78.4% after, Claude Mythos Preview 79.6% before and 85.4% after.

Download image

Before and after

Claude Opus 4.7

Before Introducing Claude Opus 4.7, 16 Apr 2026
Before: Introducing Claude Opus 4.7, 16 Apr 2026
After Introducing Claude Opus 4.8, 28 May 2026
After: Introducing Claude Opus 4.8, 28 May 2026

Claude Opus 4.6

Before Claude Opus 4.6, 5 Feb 2026
Before: Claude Opus 4.6, 5 Feb 2026
After Claude Opus 4.8 System Card, 28 May 2026
After: Claude Opus 4.8 System Card, 28 May 2026

Claude Sonnet 4.6

Before Introducing Claude Sonnet 4.6, 17 Feb 2026
Before: Introducing Claude Sonnet 4.6, 17 Feb 2026
After Claude Opus 4.8 System Card, 28 May 2026
After: Claude Opus 4.8 System Card, 28 May 2026

Claude Mythos Preview

Before Claude Mythos Preview System Card, 7 Apr 2026
Before: Claude Mythos Preview System Card, 7 Apr 2026
After Claude Fable 5 and Claude Mythos 5 System Card, 9 Jun 2026
After: Claude Fable 5 and Claude Mythos 5 System Card, 9 Jun 2026

All benchmark corrections