Claude Code adds maxEffortLevel; 2.1.266 reverts a gateway bug
The cap curbs runaway reasoning on every provider — Bedrock, Vertex, Foundry — and 2.1.267 adds 11 prompt-cache fixes and a fresh-system-prompt flag.
Copy markdown
One setting caps effort on every provider
maxEffortLevel sets a ceiling on reasoning effort across every provider, including Bedrock, Vertex and Foundry — set it at the top level or per-model under modelSettings. You can still choose a lower level by hand, but nothing blows past the cap, so a stray 'high' can't quietly run up your tokens and latency.
On 2.1.265? Update now
2.1.266 reverts a regression where CLAUDE_CODE_USE_GATEWAY started forcing Cloud-gateway sign-in on its own, breaking anyone who authenticates with API keys or custom auth headers. The variable is ignored again unless you explicitly opt in — so if your auth broke after the last update, this is the fix.
Eleven prompt-cache bugs, squashed
2.1.267 fixes 11 cache-invalidation bugs that were quietly costing money: MCP servers reconnecting caused cache misses, resumed sessions dropped extended thinking and earlier tool results, and switching models forced a full re-send of tool definitions. Expect cache hits to hold far more often.
A flag for iterating on your system prompt
--system-prompt-snapshot off renders the system prompt fresh on every request instead of reusing the one recorded when the conversation started, so edits to your prompt text take effect mid-session — handy when you're tuning an agent. The same release fixes VS Code extension-host CPU hangs and screenshot paste on WSL2.
Elsewhere: MiniCPM5-2B brings agents to the edge
OpenBMB's new 2.52B dense model averages 53.9 across 34 benchmarks — roughly Gemma 4 12B territory at a fraction of the size — and is built to run on-device with tool-calling. Weights are open on Hugging Face, putting a phone- or laptop-resident agent a weekend away.