Claude Opus 5.5 (1M-context) Text Model

Overview

Claude Opus 5.5 is a Claude model from Anthropic built for long-running agentic coding and knowledge work. It takes text plus optional visual inputs, then produces a text response. It uses always-on adaptive reasoning, so it can plan, check work, and recover from errors during multi-step tasks. This makes long sessions easier to run without losing context or coherence.

What you can build

• Codebase work over long threads, like feature implementation, refactors, and debugging plans
• Code review notes that point to specific files, functions, and failure modes
• Technical writing from source material, like design docs, specs, and changelogs
• Document Q&A from PDFs, including extraction of key claims and structured summaries
• Visual Q&A from screenshots, charts, and UI captures, with explanations in plain language
• Multi-step research briefs where you want a single model to keep state for many turns

Inputs

• A written request in plain text. You can send a single instruction, or a full chat-style message list.
• Optional attachments as files. This model accepts PDF documents and these image types: JPEG, PNG, GIF, and WebP.
• Optional system-level guidance that sets rules for tone, format, or domain constraints.
• Optional conversation identifiers you provide to keep chat history grouped across runs.
• An optional reasoning depth control. Higher settings make the model think more before answering.
• An optional cap on response length, up to 128,000 tokens.
• Optional web search permission for runs that need up-to-date facts.
• Optional continuation data from a prior run, used to resume an ongoing task after a handoff.

Outputs

The model returns a text response that follows your instruction format. For document or image inputs, the response typically includes extracted details plus an explanation of how they map to your question. When you run it in a multi-turn workflow, the platform can also return continuation data that lets you pick up the same conversation later.

Recommended settings

• Use the default reasoning depth for most tasks. Increase it only for hard bugs or long plans.
• Set the response length cap high for agentic coding turns. The model supports up to 128,000 tokens.

Limitations

• The model’s adaptive reasoning is always on. You can’t fully disable its thinking behavior.
• This model does not accept a temperature setting. Anthropic deprecated it for this model, so reasoning depth is the control you tune instead.
• This listing returns text output only. It does not generate images, audio, or video.
• Only function tools are supported: the model can ask for function calls, and your application runs them and sends back the results. Custom free-form tools are not supported, and tool choice is limited to automatic or none.
• Some clients may not show mid-turn progress text between tool-like steps, depending on how they render responses.
• Low-quality inputs reduce reliability. Scanned PDFs, blurry screenshots, and low-resolution charts can cause extraction errors.
• Very large PDFs can exceed practical limits when combined with long chats, even with a 1M-token context.

Safety & compliance

Claude Opus 5.5 includes automated safety checks that can decline certain requests. In these cases, you may receive a normal response that indicates a refusal category instead of an answer.

• Don’t request instructions for high-risk wrongdoing, including dual-use cybersecurity misuse.
• Biology-related requests can be filtered more strictly than general education or everyday health questions.
• Avoid asking for the model’s hidden chain-of-thought or internal reasoning verbatim. Some prompts can trigger a reasoning-extraction refusal.
• Review outputs before acting on them in security, medical, or legal settings.