In short
- Claude Opus 5.5 has been the top model in Anthropic's Opus line since 22 September 2026. It costs less than Opus 5 ($4 input and $20 output per million tokens) and, according to Anthropic, scores higher on nearly all of its own tests.
- Thinking is always on and the default setting is medium. That changes your costs and, for custom software, your code.
- For many tasks Sonnet 5.5 is cheaper and fast enough. Test on your own work before you switch.
Claude Opus 5.5 is Anthropic's current Opus model, released on 22 September 2026 as the first model in the Claude 5.5 family. It is meant for long-running coding and knowledge work, such as analysing large case files or carrying out tasks across several systems. It is cheaper than its predecessor Opus 5 and, according to Anthropic, more than 30% faster. The most important thing to know before switching: thinking is always on and the default setting is lower than for Opus 5, so you need to take another look at your costs and at any software of your own.
Below you can read what is new, how well the model scores according to Anthropic, and what a team or IT department should check. All figures come from Anthropic itself, so they are claims by the maker. If you want to know how Opus 5.5 compares with ChatGPT, read ChatGPT or Claude? With GPT-6 and Opus 5.5, that is the wrong question.
What is new in Claude Opus 5.5?
The main facts from Anthropic's announcement and documentation:
- Price: $4 per million input tokens and $20 output. Opus 5 cost $5 and $25. Reading from the cache costs $0.20 instead of $0.50, which is 60% less. Fast mode, up to 2.5 times faster, costs $8 and $40.
- Total cost: according to Anthropic, it costs 40% less to run than Opus 5.
- Speed: output arrives more than 30% faster.
- Memory per conversation: 1 million tokens, roughly 555,000 words, with replies of up to 128,000 tokens.
- Available on all major platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, and without data retention if you request it.
- Knowledge up to: June 2026.
Anthropic says Opus 5.5 will not be retired before 22 September 2027 (Anthropic, models overview).
How good is Opus 5.5, according to Anthropic?
Anthropic publishes scores alongside Fable 5.1, a heavier model, and Opus 5. All figures below are from Anthropic, at maximum effort:
| Test | Opus 5.5 | Fable 5.1 | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0, coding in the terminal | 66.4% | 55.8% | 52.3% |
| GDPval-AA, tasks from a range of occupations (Elo score) | 1846 | 1735 | 1708 |
| Humanity's Last Exam, hard questions with tools | 67.7% | 65.6% | 63.6% |
Figures from Anthropic itself, from the Opus 5.5 announcement (Anthropic, 22 September 2026).
So on these tests the model scores higher than Fable 5.1, which costs more than twice as much. Read that with care. Anthropic itself writes that benchmark differences at this level are a less reliable guide to real-world differences. Mind the setting too: at the default (medium) Opus 5.5 scores lower. On CursorBench that is 52.5%, against 57.8% at maximum effort. It also reports that the model often suspects it is being tested, and that finding every flaw before release is an unsolved problem.
What does "thinking is always on" mean for your bill?
Opus 5.5 thinks on every request, and you can no longer switch that off. How much it thinks is steered with a single setting: effort, with five levels from low to max. The default is medium. For Opus 5 it was high (Anthropic, migration guide).
That has two consequences:
- Thinking is billed as output. Tasks that used to run without thinking can now produce more output tokens. The lower price per token does not always make up for that.
- A comparison with Opus 5 is only fair at the same effort. Put both on high, and the prices per token tell you more than the default settings do.
So measure the cost per task over a series of real jobs, not the price per token.
Opus 5.5, Sonnet 5.5 or Fable 5.1?
| Model | Fits | Price per million tokens (input / output) |
|---|---|---|
| Fable 5.1 | Heavy reasoning, long agent tasks | $10 / $50 |
| Opus 5.5 | Long-running coding and knowledge work | $4 / $20 |
| Sonnet 5.5 | A good balance of speed and quality | $2 / $10 |
| Haiku 5.5 | Many small tasks, fast and cheap | from $0.10 / $0.50 |
Anthropic's documentation recommends Opus 5.5 as the starting point for most workloads, and Fable 5.1 only if your tests with Opus 5.5 at higher effort still fall short. Sorting, extracting or classifying in large volumes is a job for a smaller model.
Switching from Opus 5: what should IT check?
If you only use Claude through the app, or through a tool whose connection someone else manages, there is little to do. For Claude Managed Agents, Anthropic says only the model name needs changing. If your organisation has its own software that calls the Claude API, check these points from the migration guide:
- Remove settings that now cause an error. Turning thinking off, a fixed thinking budget, and the
temperature,top_pandtop_ksettings are rejected. - No more prefilled replies. A message ending with a half-written reply from the model is rejected. Use structured outputs or instructions instead.
- Replies start with thinking blocks. Code that reads the reply at a fixed position breaks. Read blocks by type.
- Set the effort yourself and recalibrate cost and speed on your own tasks.
- Leave room for long replies. At the highest levels Anthropic advises starting with a maximum of 64,000 tokens, because the thinking counts towards it.
What should you watch for?
- Safeguards. Opus 5.5 has the same safeguards for cybersecurity and biology as Fable 5.1. As a result, most cybersecurity tasks are routed to Opus 4.8. For biology there is a separate verification programme, and Anthropic says the cyber programme will soon be extended to Opus 5.5 (Anthropic, 22 September 2026). If you do that kind of work, test beforehand whether your task gets rerouted.
- No independent measurements here. All performance figures in this article come from Anthropic. A test of your own on your own tasks is the only reliable comparison.
- Human review. Even a stronger model makes mistakes. Put an owner on every application to check the result.
How to reach a decision
Pick three recurring tasks, give Opus 5.5 and Sonnet 5.5 the same input and the same effort, and judge on correctness, usability and cost per task. With that outcome you know whether the more expensive option earns its money. Repeat it at the next release.
Sources
- Anthropic (22 September 2026). Introducing Claude Opus 5.5.
- Anthropic, documentation. Models overview, accessed 26 October 2026.
- Anthropic, documentation. Claude Opus 5.5 migration guide, accessed 26 October 2026.
