Anthropic Recommends Opus 5.5 as Starting Point for Most Work

Anthropic recommends starting most coding and knowledge work with Claude Opus 5.5 at medium effort, then adjusting verification and reasoning settings before moving to a larger model to manage cost and results.

Claude Code uses Opus 5.5 by default on Pro, Max, Ultra, Team and Enterprise plans. Anthropic assigns Sonnet 5.5 to well-defined tasks, Haiku 5.5 to high-volume work such as classification and data extraction, and Claude Fable 5.2 to demanding technical projects.

API list prices per million input and output tokens are $2 and $10 for Sonnet 5.5, $4 and $20 for Opus 5.5, and $8 and $40 for Fable 5.2. Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. For longer prompts, those rates rise to $0.50 and $2.50.

To improve a result, first let the model check its work, such as by running a test, build or automated API check. The next step is to raise the effort setting. If that does not help, try a larger model at a lower effort setting; a model that lacks necessary knowledge or capability may need to be changed.

On SWE-bench Pro, Opus 5.5 at medium effort scored about 2.5 points below its high-effort result at roughly 70% of the cost. Low effort scored about eight points lower at around one-third the cost, while xhigh effort gained about 1.4 points over high at 2.5 times the cost. A retry strategy solved about 97% of tasks for roughly $0.17 each by running tasks at low effort and retrying failures at high. Running every task at high solved 95.3% at $0.29 each.

For supporting work such as finding files or reading logs, Anthropic recommends using smaller models while reserving code changes and judgment calls for the main model. In an internal defect search spanning about 22 million tokens of source code, Opus 5.5 coordinating as many as 25 Sonnet 5.5 workers cost about one-third less than Opus running alone and scored about the same. The approach adds little when tasks depend on one another or fit within a single context window, while Sonnet 5.5’s multi-agent coding results provide related benchmark context.

Switching models during a conversation requires the new model to write the conversation into its prompt cache. Anthropic recommends switching models at a natural break, such as after using Claude Code’s /compact command or starting a new conversation with a short plan. Compare models on the same work and check the cost with /usage.