TLDRocket
22 September 2026
AI spent today pulling its numbers toward work: cheaper tokens, tighter error bars, and interfaces that make it easier to get multi-step results without babysitting modes. Anthropic rolled out Claude Opus 5.5, claiming Fable 5.1-level performance on most tasks while cutting typical running costs by 40% versus Opus 5—then sweetened the deal with lower API pricing (for example, $4 per million input tokens and $20 per million output tokens, plus cheaper cached reads). OpenAI joined the same cost-and-accuracy story with GPT-6 Sol and Luna for coding and clerical tasks, positioning them as half the price of the prior GPT-5.6 Sol/Luna series and promising fewer mistakes and errors. Even the UI story moved: Anthropic merged Claude chat and Cowork into one thread, and in head-to-head comparisons ChatGPT’s Work mode finished a research task in 1 minute 17 seconds versus 1 minute 44 seconds for Claude—Claude making up time with more context and method notes.
Read the full briefing →