Hidden model Obsidian 3 tops FrontierCode 2.0 at 46.8%, suspected to be unreleased Fable 5.5
kimmonismus · x · 2026-10-28
A hidden model dubbed "Obsidian 3" was found in Cognition's FrontierCode 2.0 data, scoring 46.8% (high effort) vs. #1 Opus 5.5's 45.5% on the 132-task agentic coding benchmark. Per the chart, its max setting hits 45.5% at $47/task, ahead of GPT-6 Astra (42.8%) and Fable 5.1 (32.4%). Since it runs in the claude-code harness used exclusively by Claude models, the author concludes it's likely the unreleased Fable 5.5, listed under hiddenModels in the official JSON but not shown on the site.
More from Models
- Users find 6.1 Terra 'lazy': it reports back every 5 minutes instead of finishing — Captain-Overboard · 2026-10-28
- Community curates Awesome GPT-6.1 Terra list: projects, prompts and independent evals — Awacer · 2026-10-28
- All tests passed, but Claude Code and a local 27B still changed how 2,672 real requests got judged — Stunning-Sherbet1853 · 2026-10-28
- Zero-training PCA injection into a 256d null space takes EmbeddingGemma-2 (577M) from #26 to #1 on MAEB audio — JeanMichelRanu · 2026-10-28
- Liquid AI's LFM3 puts browser agents, voice, and vision on enterprise-controlled devices — JosephJacks_ · 2026-10-28
- Zero-training PCA head lifts 577M EmbeddingGemma-2 to #1 on MTEB audio, beating 9B models — JeanMichelRanu · 2026-10-28