Hidden model Obsidian 3 leaked on Cognition's FrontierCode 2.0, beats Opus 5.5 at lower cost
daniel_mac8 · x · 2026-10-29
danielmac8 speculates that Cognition's FrontierCode 2.0 leaderboard accidentally leaked a hidden model, Obsidian 3 (allegedly Fable 5.5):
- It scored 46.8 at high reasoning effort, beating Opus 5.5 max (45.5) and GPT-6 Astra max (42.8) at a lower cost per task ($6-7 vs $15-20 for Opus max).
- He expects it to be a leap not in coding but in reasoning, intuition, inferring human intent, ambition, creative out-of-distribution thinking, and novel solutions.
- If it really is Fable 5.5 and Cognition already benched it, the release must be imminent — possibly within days.
- A follow-up stresses that outperforming Opus 5.5 Max for less money bodes well for cost and efficiency.
More from coding & agent
- NVIDIA's AgentBaton cuts agent API costs 65.9% with a single model handoff — dair_ai · 2026-10-30
- Modeinspect turns your codebase into an editable UI, killing the Figma handoff — thisdudelikesAI · 2026-10-30
- AlphaWeb launches on Product Hunt: hosting, checkout and payments for Claude-built sites — mhdfaran · 2026-10-30
- Microsoft's TokenEconomizer Cuts Coding Agent Costs 18–31% on SWE-Bench — unixterminal · 2026-10-30
- The counterintuitive AI take: stop building more, start perfecting one thing — DanielLockyer · 2026-10-30
- Nigerian student team wins ₦10m at Zenith Bank's Zecathon 6.0 with dispute-resolution AI agent — saheedniyi_02 · 2026-10-30