Cognition's FrontierCode 2.0 data briefly listed a hidden 'Obsidian 3' above Opus 5.5, then dropped it

A copy of Cognition's FrontierCode 2.0 results file, saved by the Wayback Machine at 19:48 UTC on October 27, lists a 26th model called 'Obsidian 3' that is not in Cognition's launch post. Its best score, 46.8% at high effort, is above the 45.5% that Claude Opus 5.5 posts at max effort, the top official result. When we read the same file at 13:00 UTC on October 28, it listed 25 models and Obsidian 3 was gone. The row was first spotted by @tensorson on X.

WHAT THE ROW SAYS

Obsidian 3 was run in the claude-code harness, the one Cognition uses for every Claude model, at five effort levels. Scores: 39.2% low, 39.8% medium, 46.8% high, 45.9% xhigh and 45.5% max. Mean cost per task was $8.62 at high and $47.44 at max, against $12.82 for Opus 5.5 at max, about 3.7 times as much.

THE OFFICIAL RESULTS

Cognition published FrontierCode 2.0 on October 27: 132 tasks built with open-source maintainers, graded on whether the maintainer would merge the change. Its file ranks Claude Opus 5.5 first at 45.5%, then GPT-6 Astra at 42.8%, Claude Sonnet 5.5 at 41.0% and GPT-6.1 Sol at 39.0%. Cognition says the scores are not comparable with FrontierCode 1.1, where Opus 5.5 scored 54.6%.

WHO MAKES IT

@kimmonismus says it is very likely Fable 5.5 because of the claude-code harness. That is his reading. The file itself leaves it open: it draws Obsidian 3 in the same blue it uses for Muse Wonder 1.0, not the orange it gives every Claude model, and Claude Code can be pointed at other labs' Anthropic-compatible APIs.

WHAT WE DO NOT KNOW

Which lab made Obsidian 3, whether it is a released model, and why Cognition removed the row. Cognition has not commented, and Anthropic has announced no model by that name.