Obsidian 3 Leak Revives Fable 5.5 Rumor

// 1h agoBENCHMARK RESULT

Obsidian 3 Leak Revives Fable 5.5 Rumor

A hidden “Obsidian 3” entry in Cognition’s FrontierCode 2.0 data reportedly scored 46.8%, narrowly beating Claude Opus 5.5’s 45.5%, before disappearing. Its Claude Code harness has fueled speculation that it is an unreleased Claude Fable 5.5 checkpoint, but Anthropic has not confirmed the model.

// ANALYSIS

The leak is intriguing benchmark archaeology, not a launch announcement. The strongest clue is the harness, but the evidence remains circumstantial and the removed row could represent an internal or partner model.

–Obsidian 3 reportedly scored 46.8% at high effort versus Opus 5.5’s 45.5% at maximum effort.
–The model appeared in Claude Code, suggesting an Anthropic connection but not proving Anthropic ownership.
–Cognition removed the row, leaving its provenance, availability, and evaluation conditions unclear.
–Developers should treat the result as an unverified signal rather than evidence of a production-ready Fable upgrade.
–If genuine, Fable 5.5 could reset expectations for long-running coding agents and frontier model pricing.

// TAGS

claude-fable-5-5llmbenchmarkevaluationcoding-agentai-coding

DISCOVERED

1h ago

2026-10-29

PUBLISHED

1h ago

2026-10-29

RELEVANCE

8/ 10

AUTHOR

WorldofAI

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE26m ago

Shopify Opens Billion-Product Catalog to Agents

Shopify’s Fall ’26 Edition lets AI agents search more than one billion products through Shop.md, complete purchases with Shop Pay, and interact with every Liquid storefront through structured WebMCP tools. Australian merchants also get early access to Shop Pay Installments.

RESEARCH1h ago

WorkFactory maps intents into executable agent tasks

WorkFactory is a framework for generating coherent professional tasks and expert trajectories for training work-oriented AI agents. Its authors built 112,686 tasks across 612 occupations and fine-tuned the WorkPilot model family, with the largest model improving substantially on professional-work benchmarks.

RESEARCH1h ago

TCM-DiagCommit Reframes TCM Diagnosis Benchmarks

TCM-DiagCommit introduces a benchmark for testing evidence-conditioned syndrome selection, staged diagnostic updates, and commitment under uncertainty. Across five zero-shot LLMs, accuracy fell from 63.2–66.4% on random candidates to 14.8–22.3% across all 148 labels, while wrong-terminal commitments reached 62.2–74.3%.