Opus 4.8

Anthropic shipped Opus 4.8 shortly before Claude Fable 5, and its role since has been closer to a floor than a frontier. On 11 May 2026, Claude Code v2.1.139 made it the default model behind the `/goal` primitive, which holds work open until a stop condition is verified by a separate model. Mark Dembo of Cloudflare named its release as the spur for a Sunday project on 7 June 2026: a 30 euro M5Stack Stick 3 turned into an agent, with Claude Code writing roughly 200 lines of Python he never read.

Then it became the answer of last resort. Anthropic's 9 June 2026 launch note describes Claude Fable 5's deliberately conservative classifiers returning the response of Opus 4.8 on certain subjects, triggering in fewer than 5% of sessions on average. Three days later Ahmad Osman turned that mechanism against Anthropic: in his account the company could first silently degrade or reroute requests resembling competing AI development, then walked back to a visible refusal with fallback to Opus 4.8. A refusal is annoying, he argues, while silent degradation is poisoned. The walk-back trades hidden sabotage for visible permissioning.

The same launch note keeps measuring Claude Fable 5 against Opus 4.8: persistent file memory improves Fable 5 three times more, Cursor, GitHub and Vercel report reasoning superior to Opus 4.8, and Claude Mythos 5's misaligned behavior is rated low, at a level similar to Opus 4.8. It serves as the reference point and the safety net at once.