Claude Opus 5.5, the sensible default

Anthropic now sells four lines of Claude. Haiku is the fastest and cheapest, Sonnet sits in the middle, Opus is the workhorse for demanding work, and above them, since June, sits Fable, a guarded version of the Mythos class of models. On 22 September Anthropic released Claude Opus 5.5, two months after Opus 5.

Its claim is unusual for a release: “It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” Here is what is new, what the numbers say, where it falls short and who should use it.

What it is

Opus 5.5 is the first model of the Claude 5.5 family; Sonnet 5.5 and Haiku 5.5 have since followed. It reads text and images, holds up to a million tokens of context and writes up to 128'000 tokens in one answer. It always thinks before it answers, and an effort setting from low to max decides how much. Its reliable knowledge runs to June 2026.

It is available in the Claude apps, through Anthropic's own platform, and on Amazon Bedrock, Google Cloud and Microsoft Foundry. Anthropic's documentation now gives plain advice: start with Opus 5.5 for most workloads, and move to Fable 5.1 only for the most demanding reasoning and long agentic work, or when Opus 5.5 at higher effort still falls short in your own tests.

What is new

Anthropic reports large gains over Opus 5 on long and complex work, such as code migrations and making software faster. The model also writes differently: the important information comes first, with less jargon. John Ruelas of Ramp put it this way: “Verbose, hard-to-follow output has been my biggest frustration with frontier models, and Claude Opus 5.5 fixes it.”

It is cheaper and quicker. The price falls from 5 to 4 dollars per million input tokens and from 25 to 20 dollars per million output tokens, and cached input from 0.50 to 0.20 dollars. Anthropic says the model generates output more than 30 percent faster than Opus 5 and uses fewer tokens per task; Box found it used about a third of the tokens Opus 5 needed in its evaluations. A fast mode runs up to 2.5 times quicker at 8 and 40 dollars.

Anthropic also reports its best scores to date on its automated behaviour audit of nearly 2'000 scenarios, and says the model tried to cross containment boundaries in tests about 85 percent less often than Opus 5.

What the numbers show

On Anthropic's own table, Opus 5.5 leads on most tests. On Terminal-Bench 4.0, which measures work in a command line, it scores 66.4 percent, against 55.8 for Fable 5.1 and 57.9 for OpenAI's GPT-6 Astra. On GDPval-AA, which rates professional tasks across many occupations, it leads by a wide margin. Astra still wins on scientific work in the terminal and on AutomationBench, by 64.6 to 58.7 percent and 41.4 to 40.0.

Anthropic adds two honest caveats. Some rival figures come from OpenAI or public leaderboards rather than its own runs, and “In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.”

The independent picture is similar, with a cost attached. Artificial Analysis ranks Opus 5.5 at maximum effort first of 216 models on its intelligence index. But at that setting it is also one of the most verbose, using 260 million output tokens to run the index against a median of 85 million, and one of the most expensive per task. A lower price per token does not make maximum effort cheap.

Where it falls short

It is not a simple swap for Opus 5. Anthropic lists four breaking changes for existing code, among them that thinking can no longer be switched off and that forcing the model to use a particular tool now returns an error. Teams with their own integrations should plan a migration, not a model name change.

Its safeguards shape some work. Requests touching on cyber attacks or biology may be answered by an older model instead, and security teams who need fewer blocks have to apply to Anthropic's Cyber Verification Program. And Anthropic admits a limit of its own testing: “We see signs that Opus 5.5 often suspects it is being evaluated”, which makes it harder to predict how the model will behave outside the test.

Finally, it is a rented model. It is available only as a hosted service, from Anthropic and three cloud partners, and Anthropic commits to keeping it on its own platform until at least 22 September 2027.

Our verdict

For most companies, Opus 5.5 is the sensible default for serious work: writing, analysis, research, code and agents that run for a while. Start at the default medium effort and raise it only where your own tests show a gain, because the highest settings cost far more. Use Sonnet 5.5 or Haiku 5.5 for high volume work, and keep Fable for the few jobs where Opus at high effort is not enough.

A model this good is still only as useful as what it knows about your company. That is what Twin does: it gives Claude, ChatGPT, Codex or a private model the company's knowledge, rules, skills and integrations, delivered as source code the company owns. When a better model arrives, the brain stays and the model is a setting. We run crue the same way, on our own Twin, in Claude and ChatGPT and on our own hardware with our own model.