Claude Opus 5.5: what it is, what it changes and who it suits

Claude Opus 5.5 is the model that Anthropic released on September 22, 2026, the first of the new 5.5 family. The news is not that it is the most powerful ever — on this Anthropic itself is prudent — but that it brings the level of Claude Fable 5.1 at a 40% lower cost to that of Opus 5, with a 30% faster text generation.

For those who use AI in the company, something concrete changes: the top model begins to cost like a mid-range model from a year ago. only verified data on the official announcement and benchmark table published by Anthropic, including tests where Opus 5.5 is not first. Verified data 23 September 2026.

What has changed compared to Opus 5?

Three things, in order of weight for those who use it.

The long work. Anthropic describes Opus 5.5 as a model capable of carrying out tasks in multiple steps for hours with little supervision. In practice: you give him a whole goal in a message, you tell him what "finished" means, and he proceeds without asking you for confirmation at every step.

The automatic reasoning. Opus 5.5 always thinks before answering and decides for itself how much. The old instructions like "reason step by step" are no longer needed: removing them reduces the wait without losing quality. The detail is in Opus 5.5 Reasoning Levels.

Reading long images and documents. It reads graphs, patterns, and slides more accurately than Opus 5, especially when the meaning depends on where things in the image. And in the long documents he finds contradictions: a date that falls on the wrong day, a graph that does not come back with the numbers of the table.

How much does Claude Opus 5.5 cost?

The −40% in operating cost is not a miscalculation relative to the −20% in the price list: Anthropic says the model consumes fewer tokens to finish the same task, and the two effects add up. The actual impact depends on how your system is built; for the calculation based on your volumes, see price per million tokens, explained.

The drop in the cache is the one that moves systems to production the most, where the system prompt is sent back identical to each call. Always check the price list on the official page: it changes often.

Benchmarks, in two lines

In Anthropic's published table, Opus 5.5 leads in code and terminal tasks (Terminal-Bench 66.4%), professional work (GDPval-AA 1846 Elo) and computer use. GPT-6 Astra remains ahead on AutomationBench — by one and a half points — and on Terminal-Bench-Science. One independent result: on Artificial Analysis' Intelligence Index, it scores 58 against Fable 5.1's 53.

The complete tables, methodological notes — Terminal-Bench at different stress levels between Opus and Astra — and the line-by-line comparison are in Claude Opus 5.5: benchmarks, prices and comparison. Here the point is another: What changes for those who use it.

Opus 5.5 or Fable 5.1: Which one to choose?

Anthropic holds Fable 5.1 as a model for the most difficult work and places Opus 5.5 just below, "at the same level on most tasks". The useful question is not which is better, but for which task the difference is seen.

Practical rule: start from Opus 5.5, measure out of fifty real cases, and go up to Fable only where the numbers ask for it. The comparison between the Claude bands, with the costs, is in Claude models in comparison.

What hasn't changed

• It's not a brand new model: it is an improvement of the Opus line, and Anthropic places it under Fable for the most difficult job.
• Does not eliminate human control: on long tasks it remains necessary to define what can and cannot be done, as for any AI agent in the company.
• Safety filters are more stringent, not less. It is the first Opus with bio and cyber protections at the Fable level, and this produces some false positives on legitimate work: when a message is reported, the app switches to an earlier model. What to do in that case is in Claude moved on to an earlier model: why.

Where is it used

In the plans Claude Pro, Max, Team and Enterprise, via API with identifier claude-opus-5-5, and on Amazon Bedrock, Google Cloud Vertex and Microsoft Azure. Fast mode is available in Claude Code and on the developer platform, preview: details in Fast mode of Opus 5.5.

A novelty that affects everyone in Europe: the text generated brings a watermark (watermark) provided by the AI Act. It does not change the content, does not add tokens, does not cost anything. The framework of obligations is in AI Act: Business Guide.

Practical examples

Consulting firm, 8 people, Bergamo. He used Opus 5 to rewrite technical reports from consultants' notes, about 300 documents a month. Turning to Opus 5.5 the cost per document drops for two reasons added up: lower list and less tokens consumed. The real job of migration was to remove from prompts the lines "reason carefully before writing", which with the new model only slow down.

Commercial company, 25 employees, Padua. He wanted automatic control over quotes before they were sent: dates, totals, product names consistent across pages. It is exactly the kind of task that Opus 5.5 has improved upon —finding contradictions in long documents —and the prompt is one line: "check that there are no contradicting numbers, dates or names in this document; quote every problem and tell me where it is." The light model was not enough, the leading one before cost too much to do on every quote.

Mistakes to Avoid

• Passing everything to Opus 5.5 because it came out. To sort mail or rank requests just a light model, which costs ten times less: see the Mixed Model Strategy.
• Keeping the old prompts. Instructions on reasoning should be removed, not updated.
• Reading benchmarks as a ranking. Count your task, measured on your data, not the average score.
• Ignoring reported messages. If the app switches to a previous model without you noticing, the quality changes silently.
• Activating fast mode anywhere. It costs twice as much: it only makes sense where someone waits in front of the screen.

How to apply it in the company

1. Locate where you use a top model today and how much you spend each month. There, the −40% is immediate.
2. Locate where you are using an average model today because the top model was too expensive. These are the cases where Opus 5.5 opens up new possibilities: long contracts, documents to check, images to read.
3. Parallel Test on the same fifty cases, with the same criterion of correctness, before changing.
4. Clean up prompts from instructions on reasoning and tricks designed for earlier models.
5. Leave the light models where they are enough. The real savings comes from using the right model for each task.

Conclusion

Claude Opus 5.5 is not the news of the strongest model: it is the news of the flagship model that costs as an average. For an SME it means that automations excluded by budget a year ago today are in the accounts — as long as the process is already described, measured and verifiable, because the model remains the last cost item. The first is to connect it to what you already have.

If you want to go from testing with ChatGPT to an AI system integrated into business processes, Giallo Studio custom designed agents and automations.