Claude Opus 5.5

Frontier models are increasingly used as agents rather than chatbots. In an agent loop, one user request can trigger many model turns, tool calls, context re-reads, retries, verification steps, and long outputs. That makes two properties unusually important: how much useful work the model completes per task, and how many tokens and steps it consumes while doing it.

Anthropic positions Opus 5.5 around that production problem. The model is designed for long-running agentic coding and knowledge work, with a 1M-token context window and 128K maximum output. The release reports lower token prices than Opus 5, cheaper cache reads, faster generation, and fewer tokens per task. The core problem is therefore not simply “make the model smarter”; it is “make frontier-level agentic work economically and operationally practical.”

At the API level, Opus 5.5 is a reasoning-capable model whose thinking depth is controlled through an effort parameter; adaptive thinking is always enabled and cannot be switched off. The default effort is medium. Developers can place large repositories, documents, or multi-stage task state into a 1M-token context window and receive up to 128K output tokens in the standard API.

The agent loop can be viewed as: plan/reason -> call tools -> inspect results -> revise -> verify -> continue. Opus 5.5 is explicitly tuned for this kind of long-running work. Anthropic also documents prompt caching and a $0.20/MTok cache-read price, which can materially reduce repeated-context costs.

There are API migration details that matter in production. Thinking cannot be disabled; forced tool use returns an error; thinking blocks are tied to the model and conversation; and older computer-use tool version computer_20251124 is not accepted on the Claude API and Google Cloud. Applications that stream text between tool calls also need to account for the changed thinking-block response shape.

There are two different architectures to separate.

Model architecture: Anthropic does not disclose enough in the public release to describe the transformer topology, parameter count, routing, training mixture, or serving stack without speculation.

Production architecture: the documented system exposes a 1M context window, adaptive thinking, an effort control, tool calling, prompt caching, and deployment safeguards. For an agent product, the practical architecture is therefore a control loop around the model: application state -> model -> tools -> observations -> model -> verification -> final result, with caching and policy/safety checks surrounding the loop.

Anthropic says Opus 5.5 was evaluated with production safeguards enabled. In sensitive areas, safeguards can intervene and route work to other Opus models, which is important when interpreting published benchmark numbers.

Anthropic’s live Opus 5.5 launch page exposes several separate benchmark views. MachinoAI preserves the main comparison matrix exactly as reported, plus the pricing table and the underlying data/conditions for the Accuracy-vs-Cost charts.

MAIN BENCHMARK MATRIX

EVALUATION CONDITIONS

Unless otherwise noted, Anthropic reports Opus 5.5 with adaptive thinking at max effort. Terminal-Bench 4.0 uses xhigh effort for Opus 5.5 and high effort for GPT-6 Astra. Opus 5.5 was evaluated with production safeguards enabled; when safeguards intervened, some cybersecurity tasks were completed by Opus 4.8 and some biology/frontier-LLM-development tasks by Opus 5.

FOOTNOTES

¹ Terminal-Bench 4.0: standard error ±2.6 points for Opus 5.5 and ±1.6–2 points for the other Claude models. Anthropic’s setup reproduced Opus 5 at 52.3%; the public leaderboard reports 51.8%. ² AutomationBench: Zapier ran the evaluation without fallback models, so safeguard interventions counted as failures. ³ Terminal-Bench-Science 0.1: standard error ±3.5–5 points per model; Anthropic’s setup reproduced Opus 5 at 29.0% while the public leaderboard reports 30.0%.

PRICING — PER 1M TOKENS

FAST MODE

Up to 2.5x speed; $8 per million input tokens and $40 per million output tokens.

ACCURACY-VS-COST VISUALS

Anthropic separately publishes Accuracy-vs-Cost charts for Terminal-Bench 4.0, FrontierCode v1.1 Main, CursorBench 4.0, GDPval-AA v2.1, AutomationBench, and WANDR. These charts plot benchmark quality against task cost and change the evaluation question from raw score to score-plus-economics. Anthropic states that some default-effort Opus 5.5 results differ from the headline max-effort matrix: FrontierCode default effort is reported at 54.6%, and CursorBench default effort at 52.5%. The charts should therefore not be merged into the headline matrix as though they were the same evaluation condition.

WANDR

Perplexity’s WANDR benchmark measures large data-collection tasks. Anthropic reports Claude models under a common setup using offline web search/web fetch, programmatic tool calling, code execution, and a 980k-token task budget; Anthropic explicitly notes these conditions differ from Perplexity’s published setup and are not directly comparable across setups.