Claude Opus 5.5

The General Frontier Tier Before Fable Escalation

Opus 5.5 occupies a useful middle position in Anthropic’s current lineup. Sonnet 5.5 is cheaper and faster, while Fable 5.1 is more expensive and reserved for demanding reasoning and long-horizon work that still falls short on Opus.

That makes Opus 5.5 the practical frontier baseline: strong enough for difficult coding, code review, research, and knowledge work without immediately paying Fable pricing.

The model provides a 1,000,000-token context window, supports up to 128,000 output tokens, accepts text and images, and uses adaptive thinking on every request. Its default effort is medium, and Anthropic classifies comparative latency as moderate.

Anthropic released Claude Opus 5.5 on September 22, 2026 and lists it as Active (latest), with retirement not sooner than September 22, 2027.

• Current Opus model and recommended starting point for most workloads.
• 1M context window and 128K maximum output.
• Text and image input with text output.
• Adaptive thinking is always on.
• Lower standard token pricing than Claude Opus 5.

Accuracy Matters More Than Fluent Presentation

Research and professional knowledge work often fail in subtle ways. A model can produce a polished answer while using the wrong number, attributing a claim to the wrong source, or carrying an early assumption through the rest of the analysis.

Anthropic’s prompting guidance says Opus 5.5 is less likely than Opus 5 to state an incorrect figure or cite the wrong source. This makes it a relevant candidate for research synthesis, financial and operational analysis, policy review, document-heavy decision support, and other workflows where factual slips are costly.

The 1M context window allows a large body of source material to remain available during the task, but evaluation should still measure retrieval and attribution quality. A model that can technically receive a million tokens is not automatically guaranteed to use every relevant detail correctly.

For high-value knowledge work, score source fidelity, numerical accuracy, contradiction handling, unsupported claims, and whether the final output makes it easy for a reviewer to verify important conclusions.

Large Enough for Repository, Research, and Agent State

A 1M-token working context can combine source code, long documents, conversation history, tool results, system instructions, images, and intermediate findings without immediately forcing aggressive truncation.

This is especially useful for long-running agents because context can accumulate over many steps. It is also relevant to research and review tasks where the model needs to compare evidence distributed across large source collections.

Anthropic additionally offers a beta path for up to 300,000 output tokens through the Message Batches API using the documented output-300k-2026-03-24 beta header. That capacity is specific to batch processing and should not be confused with the normal 128K maximum output.

Large capacity should still be managed deliberately. Prompt caching, context selection, retrieval quality, and clear separation of instructions from evidence remain important even when the technical ceiling is high.

• Context window: 1,000,000 tokens.
• Standard maximum output: 128,000 tokens.
• Batch API maximum output: up to 300,000 tokens in beta.
• Minimum cacheable prompt length: 512 tokens.

Four Breaking Changes Need Explicit Testing

First, thinking cannot be disabled. Opus 5.5 uses adaptive thinking on every request, so integrations should control depth through effort rather than old thinking modes or token budgets.

Second, forced tool choice is no longer accepted. Applications using tool_choice values that force any tool or a named tool need to move to automatic selection together with strict tool use or structured outputs where appropriate.

Third, thinking blocks are tied to the model and conversation. Routers, fallbacks, or edited histories need to account for thinking-block compatibility when traffic moves between models.

Fourth, on the Claude API and Google Cloud, Opus 5.5 requires the newer computer_toolset_20260801 computer-use interface instead of the earlier computer_20251124 tool.

There is also a response-shape change: text between tool calls may appear in thinking blocks whose text is empty at the default display setting. Applications that previously surfaced that text as progress updates should test streaming behavior explicitly.