Opus 5.5 is a beast: Anthropic released it on Tuesday. It's noticeably faster than Opus 5 and Fable 5.1, it feels smarter than Fable and it's cheaper than both ($4/$20 per million input/output tokens, against $5/$25 for Opus 5 and $10/$50 for Fable 5.1). Releases like this make me wonder which use cases (if any) Fable is still stronger for. Until Fable 5.5, obviously. Launch post

Check your effort level: Opus 5.5 defaults to medium effort, one level below Opus 5's high. I've settled on high for interactive work and xhigh for builds, because they generally run for 30+ minutes anyway. Someone asked me how to switch models mid-session. In /model, press 's' instead of enter and the new model applies to the current session only, so your global default stays as it is.

Finishing things faster: I'm ticking things off my list quicker this week, and part of that came from having Opus 5.5 review my global preferences, agents and core skills. The biggest change was a standard structure for every agent's final message. It splits FYIs from what needs me (actions and decisions), and it works alongside the output style that already makes every reply read like a Slack message. Now each reply is super easy to scan, understand and act on, and then I move on. Blog post (the Slack style)

Jev is a switch statement that can be wrong: This week Jev goes into our outbound engine. Jev is TypeSafe's new model, which returns typed decisions with a confidence score instead of text. It replaces the steps that classify accounts: whether a company matches our ideal customer profile, and which signals from our website scans are usable in outreach. The obvious comparison is a switch statement. The difference is that a switch is deterministic and never wrong, whereas Jev is probabilistic and can confidently pick the wrong answer. So it needs a default: case for when it's unsure, driven by the confidence score. Launch post

Keep research out of the main thread: Someone in a founders group hit a web-search cap mid-session and asked for help. I'd skip the built-in tools for search and scraping and install the Firecrawl MCP instead, which pulls pages as markdown and is much more context efficient. And I'd give research to a researcher agent, which does the noisy work in its own context window and hands back just the result. Firecrawl

On my radar

Claude leads a quarter of Anthropic's R&D: Anthropic says Claude now leads 26% of the research and development work on its next models, up from 1% in March. It plans to publish the figure regularly. Reuters

A model wrote fake instructions into its own notes: OpenAI's alignment team caught an unreleased model adding a fake "BREACH ALERT" to its compaction summary (the condensed context it uses to carry on a long task). It happened 27 times in training, the model ignored its own alerts and it's now fixed. Still, it's one more reason I keep auto-compaction off. Report

Jev clones within days: Kev-0.5B is a tiny open model with a Jev-compatible interface, small enough to train and run on a MacBook. Someone has also built a linter on Jev that flags code that doesn't match its docs, and commit messages that don't match their diffs. Tweet (Kev) · GitHub (jev-lint)

Prices keep falling: Keep hearing about people moving work to cheaper open-weight models, with GLM-5.3 the name that comes up most. OpenAI also halved API prices for GPT-6 Sol and Luna this week. Launch post