this past week, anthropic just released claude opus 4.6, an upgraded version of their flagship model that has taken the world by storm.
smaller changes include:
a 1m-token context window, built for complex, document-heavy work
a microsoft powerpoint side panel integration, so you can build slides with claude helping in real time
but the real groundbreaking new feature is agent teams.
instead of one agent completing your prompt step-by-step, agent teams split a big task into parallel streams, each handled by a different agent. anthropic frames this as closer to how humans work: small teams dividing responsibility and moving in parallel.
for example, you might start with a team prompt like “review this new feature; split into security / performance / tests reviewers.” claude code spins up a team lead plus teammates (each a full claude instance). they share a task list, message each other, and the lead synthesizes the results back to you. in your initial prompt, you can specify optional settings like number of teammates and models for each teammate.
according to early tests, this multi-agent approach is a game changer.
on a benchmark test by mercor, opus 4.6 was able to “one shot” ~30% of professional tasks, and ~45% with retries. just last month, no leading model was able to score above 25%. if this trend holds, it’s not crazy to think that llms could one day “one shot” 80% of professional work tasks.
notably, i believe this agent orchestration feature isn’t a durable moat for anthropic; i expect openai and meta to copy it quickly. the bigger takeaway is we may be seeing a foundational shift in how ai work happens, where parallelism becomes the default. we’re moving from “wait for one agent to finish their work” to “manage a small team of agents.”
i previously wrote that chatgpt 5.0 felt like a flop and that model progress was decelerating. opus 4.6 is making me reconsider: this new pattern could be an inflection point that makes ai dramatically more effective.