Mystery Model "Obsidian 3" Surfaces in Cognition FrontierCode Data, Scores 46.8% Over Opus 5.5
fekdaoui · x · 2026-10-28
Screenshots of Cognition's FrontierCode 2.0 data appear to reveal a hidden model called "Obsidian 3", scoring 46.8% at high effort versus 45.5% for the top-ranked Opus 5.5 at max. Notably, Obsidian 3 dropped to 45.5% at max effort — more inference spend didn't help on this run. It ran in the claude-code harness, which only Claude models use, fueling speculation it's "Fable 5.5"; identity is unconfirmed.
The accompanying chart adds key caveats: values are screenshot-reported and unverified, the small deltas (46.8%/45.9%/45.5%) don't establish statistical significance, and the model identity is unconfirmed.
FrontierCode grades engineering requirements across 132 tasks, including things a developer should infer even when unstated — fixing the immediate bug can still leave a change not ready to merge.
More from Models
- User Slams ChatGPT for Failing to Read Images, Jokes It Feels Like a GLM Model — DevDminGod · 2026-10-28
- Qwen's dark 'if no one is watching' rant screenshot becomes instant AI meme — teortaxesTex · 2026-10-28
- Economist Jason Furman maxes out 3 of 5 Claude accounts, told to switch to Codex — soumitrashukla9 · 2026-10-28
- Gemma 4 reportedly boosts multilingual training mix, but German still lags English — MaziyarPanahi · 2026-10-28
- Anthropic Ditches Coding Interviews as Genie 4 Launches and AI Claims ETH Disproof — Latent Space · 2026-10-28
- Rumor: gemini-omni-1.2-flash already deployed, release imminent — lyraxana · 2026-10-28