On Wednesday, Anthropic released Opus 4.6. Thirty-five minutes later, OpenAI dropped GPT-5.3 Codex. Neither company acknowledged the other. Everyone else spent the day running benchmarks and arguing on Hacker News.
Two things from the week actually matter for how you use Claude Code. Let me cut to those.
Opus 4.6: what you need to know
The launch post is long. Here's what changes your day-to-day:
1M token context window (beta, API only). 5x jump from 200K. It actually remembers what it read at the beginning now.
128K output tokens. Doubled. Fewer truncated files.
Fast mode. 2.5x faster output in Claude Code. Available in v2.1.36+. Toggle with
/fast.Adaptive thinking. The model decides when to use extended reasoning. You no longer toggle it manually.
Skip the benchmarks. The thing I am noticing: it feels better at multi-step tasks and can run much longer without babysitting.
Agent teams change the game
This is the big one. Agent orchestration was always going to be the next big thing.
Agent teams let a lead Claude Code session spin up multiple teammate agents that work in parallel. Each gets its own context window and can message other teammates directly. It's not subagents - subagents run a task and report back. Teammates collaborate.
Try it: Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in your environment. Research preview — rough edges expected.
Where it works well: parallel code reviews (one agent per module), cross-layer features (frontend + backend + test agents), debugging with multiple hypotheses simultaneously.
The catch: it eats tokens. So be mindful of what you use it for and keep an eye on it.
The C compiler demo (in brief)
Anthropic tasked 16 Opus 4.6 agents to build a C compiler from scratch. $20K in API costs. 100,000 lines of Rust. It compiles the Linux kernel on x86, ARM, and RISC-V, passes 99% of the GCC torture test suite, and can compile PostgreSQL, Redis, and Doom.
Yes, it's missing things. The point isn't that it replaces GCC. The point is 16 agents coordinated to build something non-trivial for $20K.
Workflows worth stealing
The /handover command. @zarazhangrui built a slash command that generates a handover.md when a session ends — decisions, pitfalls, lessons learned. Next session starts with full context.
Quick hits
Claude is permanently ad-free. Anthropic committed in writing: "There are many good places for advertising. A conversation with Claude is not one of them." They also ran some pretty hilarious Superbowl ads targeting OpenAI.
Memory features are live. Agent Memory and Auto Memory — Claude records and recalls memories as it works.
The memes
That's the week. A new model, a new way to coordinate agents, and a community figuring out how to not burn through their limits doing it.
If you've found a good workflow for using teams, reply and tell me. I'll share the best ones next week.
Abhishek