Summary
Theo of t3.gg gives his first impressions of Anthropic’s Opus 5.5, which he says finally fixes the problems that made recent Opus releases feel unreliable. He recalls how Opus 4.5 changed the way developers write code with AI, then explains why Opus 5 and Sonnet 5 disappointed him, citing sloppy text output and code that missed important details.
According to Anthropic, Opus 5.5 performs at the level of Fable 5.1 on most work while costing 40% less to run than Opus 5. Theo walks through Artificial Analysis data showing it on the best price-to-performance line for scores above 50, and notes that it uses about twice as many tokens but avoids a large price hike thanks to lower pricing and cheaper cache reads. He also highlights the open-weight Mimo V2.6 as a strong value option and finds Grok 4.7 and 4.6 more expensive and lower scoring than the competition.
A recurring theme is his skepticism of max reasoning effort. Medium effort on Opus 5.5 costs about $1.34 per task on Artificial Analysis versus over $6 for Fable, and a max-effort run of his own looped for six and a half hours on a single plan. He also compares the model against GPT-6 Soul and Luna on Skatebench, and the video includes a sponsor segment on the agent-authentication standard from WorkOS.
📺 Source: Theo – t3․gg · Published September 23, 2026
🏷️ Format: News Analysis
