Opus 5.5 takes the default seat in Claude Code

Claude Code 2.1.280, released on September 22, moves Pro and Team Standard plans from Sonnet to Opus as the default model, its release notes on GitHub say. The default Opus is now Opus 5.5, and an effort level saved before /effort became per-model no longer applies to newly released models, so a session on those plans that keeps the default model runs Opus 5.5 at its default effort until a level is picked.

The change matches Max, Team Premium and Enterprise plans, which defaulted to Opus before. Opus 5.5 (claude-opus-5-5) has a 1M-token context window and costs $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. With the model’s launch, Anthropic said it is raising five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and giving subscribers a rate limit reset they can save and use whenever they choose.

Anthropic puts the input and output prices 20% below those of Opus 5, which v2.1.219 made the default Opus model on July 24, and the cache-read price 60% below; cache reads, by its account, make up the majority of agentic and coding work costs. Its tests show, the company says, that at default settings Opus 5.5 will cost 40% less than Opus 5 on typical workloads, because it costs less per token and uses fewer tokens per task. Anthropic also describes the model as performing at the level of Claude Fable 5.1 on most work. Fable 5.1 was priced at $10 and $50 per million input and output tokens and $0.25 for cache reads when it became the default Fable model in September, only five cents above Opus 5.5 on cache reads.

Which effort level wins?

Opus 4.7, Opus 4.8 and Fable 5 stop holding their launch-default effort over a level set elsewhere: /effort in -p or the Agent SDK, an effortLevel in project, managed or --settings configuration, or a per-model level. Opus 4.8 arrived in v2.1.154 on May 28 with high effort by default, so for that model a level set in any of those places now takes effect.

In VS Code, the effort chip now shows the level the session runs at; it had been showing a stale saved level.

Cache misses

A model switch made from a host app (Claude Desktop, VS Code or the SDK) while Claude is working no longer causes a prompt-cache miss on the next prompt. Resumed fork subagents re-send the tool list they first used instead of rebuilding it, which had broken prompt caching for that agent. And /cost now names thinking mode and thinking display changes among the causes of a cache miss.

Auto mode and permissions

When a safety check declined to review an action, auto mode retried it over and over; the action is now denied once, with a note that retrying won’t help. A safety check that gave no answer had auto mode denying actions without pause; retries now back off, and after ten in a row the turn stops with a message.

Writes through a symlinked path were judged by their in-tree spelling. The permission prompt now names where the write lands, and acceptEdits, allow rules and auto mode no longer approve a write that lands outside the tree.

An agent-type hook no longer runs on PermissionRequest, since its answer could never allow or deny the request; in its place Claude Code shows an error pointing to command or http hooks. Plugin marketplaces whose name imitates a reserved marketplace name are refused when added, and one added earlier stops loading.

MCP text and dialog keys

CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH changes the 2,048-character cap on MCP tool descriptions and server instructions, for every MCP server in the session. The cap goes back to v2.1.84 on March 26, which set it “to prevent OpenAPI-generated servers from bloating context”.

Pressing Ctrl+C or Ctrl+D twice in most dialogs, /model and /permissions among them, now closes the dialog instead of quitting Claude Code. A stray n no longer closes a dialog and a stray y no longer confirms one: Enter accepts, Esc cancels, and binding y and n to confirm:yes and confirm:no in keybindings.json brings the old keys back.