Mystery Model 'Obsidian 3' Appeared in Cognition's Data, Then Vanished

A model that appears nowhere on the leaderboard quietly scored 1.3 points above the winner, and then disappeared. According to the reports, the results data for Cognition's newly released coding benchmark, FrontierCode 2.0, briefly contained a mystery model called Obsidian 3.

What the data showed

Cognition published FrontierCode 2.0 on October 27: 132 tasks built with open-source maintainers, graded on whether a maintainer would merge the change. On the official table, Claude Opus 5.5 ranks first at 45.5%, followed by GPT-6 Astra at 42.8% and Claude Sonnet 5.5 at 41.0%.

But ArtificialWatch reports that a web archive copy of the results file from October 27 listed a 26th model, Obsidian 3. At high effort it scored 46.8%, 1.3 points above Opus 5.5's best. At other effort levels it reportedly got 39.2% (low), 39.8% (medium), 45.9% (xhigh) and 45.5% (max). The data put its mean cost at $8.62 per task at high effort, versus $12.82 for Opus 5.5 at max. By the next day, the row was gone from the file.

Why people suspect Claude

The key clue is the harness. Obsidian 3 ran in claude-code, the environment Cognition uses for every Claude model. Some in the community take that as a sign it is the unreleased Claude Fable 5.5.

There are details pointing the other way. The file draws Obsidian 3 in the same blue as Muse Wonder 1.0, not the orange it gives Claude models, and Claude Code can in principle be pointed at other labs' Anthropic-compatible APIs. So it is a guess, and nothing official backs it.

Cognition hasn't commented on the removal, and Anthropic hasn't announced any model by that name.

What stands out

The race at the top is tight, and 1.3 points on 132 tasks isn't a big gap. The more interesting part is cost: by these numbers the model scored higher at lower spend.

Note too that Cognition says FrontierCode 2.0 scores aren't comparable with version 1.1, where Opus 5.5 scored 54.6%. Now the question is whether Anthropic or Cognition says something, or whether a model shows up on a public API that matches the name.