Claude Opus 5.5 hard-errors on disabled thinking and tool_choice:any
Migrating to Anthropic's new flagship means fresh 400s — plus ~20% cheaper tokens, always-on thinking, and a beta that adds tools without busting your cache.
Copy markdown
Three API calls that now hard-error
Point your code at `claude-opus-5-5` and three old patterns return 400. `thinking:{type:"disabled"}` is rejected — drop the field and set depth with the new effort parameter. `tool_choice` of `any` or `tool` also 400s; use `auto` with strict tool use to force a call. And computer use now needs the `computer_toolset_20260801` toolset — the older `computer_20251124` errors on the Claude API and Vertex, though it still runs on Bedrock.
Cheaper tokens, thinking you can't switch off
Opus 5.5 lists at $4/$20 per million input/output tokens — about 20% under Opus 5's $5/$25 — with a 1M-token context and 128k max output. Adaptive thinking is always on; you trade latency for depth through the effort parameter instead of toggling reasoning.
Add tools mid-conversation without nuking your cache
A new beta (`inline-tools-2026-09-15`) lets you define a tool inside a mid-conversation system message via a `tool_addition` block — so you can add, swap, or move tools without editing the top-level `tools` array or invalidating your prompt cache. MCP toolsets work the same way with the `mcp-client-2026-09-15` header.
Fast mode and cache diagnostics graduate
Fast mode (research preview) is now live for Opus 5.5 on the Claude API for lower-latency turns. Separately, cache diagnostics left beta on Sept 23: drop the old `cache-diagnosis` header and just include a `diagnostics` object on your Messages request to see per-request cache hit/miss detail.