Claude

• Claude Opus 5.5default

  claude-opus-5-5Released 2026-09Source

  Anthropic: most workloads start with Opus 5.5. Long-running agentic coding and knowledge work. Adaptive thinking always on; default effort medium; 1M context, 128K output. Default model in Claude Code on Pro, Max, Team, Enterprise and the API.

• Claude Fable 5.1top

  claude-fable-5-1Released 2026-09Source

  Most capable model open to all customers, for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short. Slower; thinking always on; API default effort high. Opt-in in Claude Code (/model fable).

• Claude Sonnet 5.5fast

  claude-sonnet-5-5Released 2026-09Source

  Faster, lower-cost complement to Opus 5.5 for well-scoped everyday tasks, bug fixing and polished documents, slides and spreadsheets. Default effort high on the API, medium in Claude Code and the apps.

• Claude Haiku 5.5budget

  claude-haiku-5-5Released 2026-10Source

  Cheapest and fastest: summaries, classification, extraction, routing and sub-agents. First Haiku with adjustable effort; default effort medium.

Where it runs

• claude.ai Projects (project instructions + project knowledge; all plans)
• claude.ai Skills (upload a ZIP with a SKILL.md folder under Customize > Skills; code execution must be on)
• Claude Code (CLAUDE.md, .claude/skills, plugins)
• Claude API (also Amazon Bedrock, Google Cloud, Microsoft Foundry)

Prompting notes

• Control depth with effort (low, medium, high, xhigh, max), not prompt wording. Opus 5.5 and Haiku 5.5 default to medium; Fable 5.1 and Sonnet 5.5 default to high on the API. Raise effort for hard steps and lower it for routine ones. Source
• Put long documents near the top of the prompt and the question and instructions at the end; Anthropic reports up to 30% better answers on complex multi-document inputs. Wrap each document in <document> tags with <source> and <document_content>. Source
• Do not ask Fable 5.x, Opus 5.x or Sonnet 5.5 to write out their reasoning in tags; such prompts may be declined. Rely on built-in adaptive thinking and lower effort if cost matters. Pre-writing the start of Claude's reply returns a 400 error since Claude 4.6: ask for the format directly instead. Source
• Use plain instructions ('Use this tool when...') instead of all-caps urgency; current models follow the system prompt closely and over-trigger on aggressive wording. Separate instructions, context and input with XML tags. Source
• Fable 5.1 writes fewer progress updates during long tool runs and uses less bold, headers and lists in chat: ask explicitly for updates and for the structure the output needs. Source