Anthropic released Claude Opus 5.5 on September 22, 2026, and priced it below the model it competes with. The model runs at $4 per million input tokens and $20 per million output tokens. That is 20% under Opus 5 and 60% under Fable 5.1, which still bills at $10 and $50. Anthropic says the model performs at the level of Fable 5.1 on most work.

OpenAI GPT-6 Astra leads in autonomous computer navigation and multi-step digital task execution, Claude Opus 5.5 targets fast, cost-efficient repository-scale software engineering, and Claude Fable 5.1 is optimized for deep reasoning, root-cause debugging, and frontier research.

What Claude Opus 5.5 pricing changes

The company also cut cache costs. Cache writes dropped 20%, from $6.25 to $5 per million tokens. Cache reads fell more sharply, from $0.50 to $0.20. On top of the token cuts, Anthropic claims a typical workload costs about 40% less than Opus 5, because the model finishes tasks in fewer tokens. The token price is fixed. That 40% figure depends on the task. Buyers weighing Claude Opus 5.5 pricing against Fable 5.1 get the same output for less on standard rates, with caching favoring longer agent runs.

Opus 5.5 vs Fable 5.1 on benchmarks

On Anthropic’s own table, the Opus 5.5 vs Fable 5.1 result shows Opus 5.5 ahead on agentic coding, knowledge work, and reasoning. Anthropic adds that the real-world gap is narrower than the scores suggest, so Fable 5.1 may still hold an edge on hard, open-ended tasks. The Claude Opus 5.5 benchmarks include a Terminal-Bench 4.0 score of 66.4%, about ten points clear of Fable 5.1 and GPT-6 Astra.

That result put the model at number one on the Artificial Analysis Intelligence Index. External evaluators, including METR and Frontier Design, tested it before release. Anthropic points to gains on long work. One tester finished a 680,000-line code migration in under a day, work the company says would take an engineering team weeks. In an internal test that cut page load times across a web app, the model succeeded 39 times out of 40. Adaptive thinking stays on by default, and an effort setting controls depth, latency, and cost in place of manual thinking budgets.

A developer test on speed

Speed is the second claim, and one hands-on test put numbers on it. A developer built a simulated trading engine with 150 bots and a live order book, then ran it on both models. Opus 5.5 finished in 12 minutes. Fable 5.1 took 42. The Opus run cost $3.61 and used 24 turns. Fable’s run cost about $17 across 62 turns. The outputs matched. A separate vendor test translating HAProxy from C to Rust showed Opus 5.5 finishing in 9.5 hours against 12 for Fable 5.1, at 51% lower cost.

Safety and rollout

Opus 5.5 is Anthropic’s first release since the company said it would pace the frontier. It scored the strongest to date on the firm’s automated behavioral audit, a suite of about 2,000 simulated scenarios. In one containment test, the model tried to cross its set boundaries roughly 85% less often than Opus 5 or Mythos 5.1. Anthropic reports better resistance to prompt injection than Opus 5. It is the first Opus model to carry the biology, cybersecurity, and AI development classifiers used on Fable 5.1, so it refuses more requests. Text outputs carry a watermark.

The Claude 5.5 family starts here. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks. Claude Opus 5.5 is available now through the Claude API and across AWS, Google Cloud, and Microsoft, and in GitHub Copilot.

Model Comparison Matrix

Deployment Guide

• Deploy Claude Opus 5.5 for software engineering teams that need continuous code reviews, large-scale migrations, and high-throughput document intelligence where speed, budget, and predictability are paramount.

• Deploy GPT-6 Astra when the objective requires hands-on desktop agency—such as operating legacy web apps, conducting end-to-end browser research, or building visual web tools directly inside a workspace.

• Deploy Claude Fable 5.1 for high-stakes problem spaces where getting to the foundational truth matters more than raw latency, including architectural refactoring, complex mathematical proofs, or multi-day scientific simulations.