On September 22, 2026, Anthropic released Claude Opus 5.5, the first model in its Claude 5.5 family. Anthropic says the model performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Claude Opus 5, thanks to lower prices and fewer tokens per task. For developers, the launch is also a migration event: the model always thinks, rejects forced tool use and ties its reasoning blocks to the conversation that produced them, so existing Opus 5 integrations may need changes before switching.

Key Facts#

  • Released: September 22, 2026, as the first Claude 5.5 model. Anthropic said Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks.
  • Availability: all platforms, including AWS, Google Cloud and Microsoft Azure. The API model ID is claude-opus-5-5.
  • Pricing: $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. Cache reads cost $0.20 per million tokens, 60% below Opus 5.
  • Speed: Anthropic says output is more than 30% faster than Opus 5. A fast mode of up to 2.5x speed costs $8 and $40 per million input and output tokens.
  • Vendor benchmarks: Anthropic reports 66.4% on Terminal-Bench 4.0 (Opus 5: 52.3%), 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0.
  • Pre-release testing: external evaluators including METR and Frontier Design tested the model before launch.
  • Safeguards: cybersecurity, biology and distillation safeguards similar to Claude Fable 5.1's, with flagged requests falling back to another model.
  • Breaking changes: thinking cannot be disabled, forced tool use returns an error, thinking blocks are bound to the model and conversation, and the older computer_20251124 tool is rejected on the Claude API and Google Cloud.

What Happened#

Anthropic positioned Opus 5.5 as a step up from Opus 5 in capability and a larger step in efficiency. Its benchmark table showed Opus 5.5 ahead of Fable 5.1 and Opus 5 on agentic coding, knowledge work and computer use. It also led OpenAI's GPT-6 Astra on the two agentic coding benchmarks with Astra scores, though it trailed Astra on AutomationBench and on a scientific research benchmark. Anthropic itself cautioned that at these capability levels "benchmark margins have become a less reliable guide to real-world differences," and all of the figures are vendor-reported.

The examples leaned toward long, sprawling jobs. In an internal test, Opus 5.5 and Fable 5.1 both translated the HAProxy load balancer from C to Rust and passed nearly all of its regression tests, but Opus 5.5 finished in 9.5 hours instead of 12, at 51% lower cost. An early tester audited and fixed a 200,000-line codebase in under three hours, a job that took Opus 5 more than 20 hours and 2.5 times as many tokens. Anthropic also said the model writes more clearly, addressing feedback about Opus 5.

Safety took a large share of the announcement. It was Anthropic's first release since CEO Dario Amodei argued that AI progress should be paced so safety practices stay ahead of capabilities. On an automated behavioral audit of nearly 2,000 scenarios, Opus 5.5 scored better than any recent Claude model on nearly every misalignment measure. In a new containment test, it tried to cross boundaries about 85% less often than Opus 5 or Claude Mythos 5.1. Anthropic also flagged a limitation: the model often suspects it is being evaluated, which makes pre-deployment testing harder to trust.

Because Anthropic rates Opus 5.5 comparable to Claude Mythos 5.1 in biology and cybersecurity, it is the first Opus model to launch with Fable-class safeguards. Routine bug finding and fixing is allowed, but most cybersecurity tasks are rerouted to Claude Opus 4.8. Vetted life sciences organizations can apply to a verification program for broader biology access. The model also ships with preserved thinking, an anti-distillation safeguard that stops API users from editing earlier context to extract its reasoning.

Background#

Opus 5.5 arrived two months after Claude Opus 5, released on July 24, 2026 at $5 and $25 per million tokens, which Anthropic said came close to Fable 5's intelligence at half the price. Opus 5 also added automatic fallbacks to the API in beta, which route requests flagged by safety classifiers to another model instead of blocking them.

Until this launch, Anthropic's strictest safeguards were reserved for its top tier, Claude Fable 5 and Mythos 5 and their 5.1 successors. Opus 5's cyber classifiers, for example, were less restrictive than Fable 5's. Opus 5.5 brings Fable-class safeguards to the Opus line, which Anthropic designs for everyday use.

Why It Matters for Developers#

Treat this upgrade as a code change, not a configuration change. Anthropic's migration notes list what breaks:

  • Thinking is always on. Requests that disable thinking or set a manual token budget return a 400 error. Use adaptive thinking and control depth with the effort parameter, which defaults to medium.
  • Forced tool use is gone. A tool_choice of any or a named tool returns a 400 error. If you forced a tool to get JSON, switch to strict tool use with automatic tool choice, or to structured outputs. Our guides to structured outputs and function calling in C# cover both patterns.
  • Keep conversations append-only. For accounts created on or after August 31, 2026, replaying a thinking block after the system prompt, tools or earlier messages have changed returns a 400 error by default. Code that rebuilds the system prompt with the current date, adds tools mid-session or trims old turns needs rework. Anthropic recommends mid-conversation system messages instead.
  • Read content blocks by type. Responses can start with thinking blocks whose text is empty by default, so code that assumes the first block is text will break.

.NET teams should check how their SDK or abstraction round-trips thinking blocks before switching model IDs. The official Anthropic NuGet package published version 12.50.0 on launch day. Run your own evaluation suite first, as described in our guide to evaluating AI applications, and track cost per task rather than per token, since Anthropic's savings claim depends on fewer tokens as well as lower prices.

What's Next#

Sonnet 5.5 and Haiku 5.5 are due within weeks. Anthropic also plans to expand its Cyber Verification Program to Opus 5.5, with three tiers of increasingly permissive access for verified security practitioners. The larger open question is Anthropic's own: how to evaluate models that increasingly recognize tests. It says progress on interpretability is needed to keep that problem from growing.

Sources#