Claude Opus 5.5: What Changed and Why This Model Feels Different

Claude Opus 5.5 is Anthropic’s September 22, 2026 flagship Opus release, positioned around long-running agents, coding, knowledge work, vision, and computer use. Its public changes include always-on adaptive thinking, a 1M-token context window, a 128K maximum output, lower pricing, cheaper cache reads, faster output, and improved communication. Anthropic does not publicly disclose the underlying parameter count or full architecture.

Opus 5.5 is presented less as a narrow benchmark release and more as a model for sustained professional work. The release emphasizes long-running coding agents, multi-tool workflows, computer use, financial analysis, and document-heavy knowledge work, while changing how reasoning configuration and cost are exposed to developers.

The model targets workloads where short-context or single-turn systems struggle: large repositories, multi-step computer tasks, dense documents, and tasks requiring repeated tool use. For these workloads, reliability depends not only on raw model capability but also on reasoning depth, context retention, latency, and the surrounding agent harness.

Opus 5.5 belongs to Anthropic’s Opus family and follows Opus 5. Anthropic describes it as operating at the level of Claude Fable 5.1 on most work while costing about 40% less than Opus 5 on typical token-billed workloads. Public materials focus on behavior and evaluation rather than exposing the transformer architecture or training mixture.

The public release is characterized through product specifications, benchmark evaluations, internal workload examples, customer reports, and safety evaluations. Anthropic evaluates the model at specified effort settings and with production safeguards enabled. The interpretation here keeps vendor-reported measurements separate from customer testimonials and independent evaluations.

Anthropic does not disclose the parameter count, transformer topology, routing design, or complete inference stack. The observable system architecture instead consists of the Claude model, adaptive-thinking control, tool interfaces, computer-use capabilities, context management, and production safeguards. Developers interact through the model ID claude-opus-5-5 and supported API/tool contracts.

The release does not introduce a single training dataset or disclose a complete training corpus. Evaluation uses multiple benchmarks, including Terminal-Bench 4.0, FrontierCode v1.1, CursorBench 4.0, GDPval-AA v2.1, AutomationBench, Humanity’s Last Exam with tools, Terminal-Bench-Science 0.1, OSWorld 2.0, and Chartography with tools.

Anthropic does not publish a full training recipe for Opus 5.5, including parameter count, complete data mixture, or optimization configuration. The public release instead exposes inference-time controls such as adaptive thinking and effort. Thinking is always enabled, with effort used to control reasoning depth, latency, and cost.

Anthropic’s comparison matrix includes Claude Fable 5.1, Claude Opus 5, GPT-6 Astra, and GPT-5.6 Sol on selected benchmarks. The release also discusses independent or third-party evaluations. Comparisons should be interpreted with their stated harness, effort, safeguard, and provider conditions rather than as a single universal ranking.

The release does not provide a conventional component-by-component ablation of the underlying model architecture. It does, however, expose meaningful evaluation dimensions: effort level, safeguard intervention, tool configuration, and workload type. These factors can materially change measured performance and should be treated as experimental conditions rather than implementation details.

The public release does not reveal the full model architecture or training recipe. Benchmark results are condition-dependent, and Anthropic explicitly notes that benchmark margins can be a less reliable guide to real-world differences at this capability level. Safeguard interventions can also affect task outcomes, and independent evaluations may produce different results.

Opus 5.5 is a frontier Opus release centered on sustained agentic work, coding, knowledge work, and computer use, with lower token and cache-read prices and always-on adaptive thinking. The important engineering change is not one disclosed architectural component but the combined model, inference controls, tool interfaces, and safeguards that determine end-to-end workload behavior.

Primary source: Anthropic, “Introducing Claude Opus 5.5,” September 22, 2026. Additional source: Anthropic Claude Opus documentation and the Claude Opus 5.5 system card.