Claude Opus 5.5 Is Out: Cheaper, Faster and Easier to Read

What was announced?

Anthropic announced Claude Opus 5.5 on 22 September 2026. It is the first model in the new Claude 5.5 family. According to the company, it performs at the level of Claude Fable 5.1 on most work. Compared with Opus 5, it costs 40% less to run on typical workloads.

The model went live on every channel on launch day. It is available in the Claude apps, on the Claude Platform API, and on Amazon Web Services, Google Cloud and Microsoft Azure. Its API name is claude-opus-5-5. Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.

The tone of the announcement stands out too. Alongside the strong results, the company is open about the limits of its own measurements. We have tried to keep that context when quoting the numbers.

What changed in performance?

In the published benchmarks, Opus 5.5 leads in three areas: software work, computer use and knowledge work. On Terminal-Bench 4.0, which measures multi-step tasks in a command line, it scored 66.4%. Opus 5 scored 52.3% on the same test. On OSWorld 2.0, which measures real tasks on a screen, the result rises from 74.0% to 81.8%. These scores were taken at the model's highest effort settings.

Knowledge work tells a similar story. On GDPval-AA v2.1, which covers real work across 44 occupations, Opus 5.5 scored 1846 Elo. Opus 5 scored 1708.

The company's own experiments are more telling than the scores. In one, the model wrote a report on a company's quarterly results using a copy of the web where the source was hard to find. Every figure and quote was checked against sources, and a single invented one failed the report. Sixteen of Opus 5.5's 18 reports passed. Neither Fable 5.1 nor Opus 5 passed in any attempt.

In another test, the widely used HAProxy software was translated from C into Rust. Opus 5.5 finished in 9.5 hours. Fable 5.1 needed 12 hours for the same job. Opus 5.5 also cost 51% less.

The company adds a caveat. At this level, benchmark margins are a less reliable guide to real-world differences. In its own use, the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest.

The real story: cost

For businesses, price may be the most important change. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. That is 20% less than Opus 5. Cache reads cost $0.20 per million tokens, down from $0.50 on Opus 5.

Cache reads look like a small line item at first. Yet they make up most of the cost of coding and agent work. The model also uses fewer tokens per task. Together, these bring the total cost down by 40% on typical workloads.

Speed is up as well. Opus 5.5 generates output more than 30% faster than Opus 5. There is also a fast mode in Claude Code and on the Claude Platform. It offers up to 2.5x speed at $8 per million input tokens and $40 per million output tokens.

Subscribers get something too. Five-hour usage limits are going up on Pro, Max, Team and seat-based Enterprise plans.

It writes more clearly

One of the most common pieces of feedback about Opus 5 was that its writing ran long and wandered. The new model is noticeably different here. It puts the most important information first. It leans less on jargon and odd phrasing. It also follows the writing rules you give it more closely.

Early testers say the same. Rogo reports shorter and better-structured answers. Clio says the code comments became short and useful instead of long and prose-heavy.

There is a safety side to this. Output that is easy to read is also easy to check. In processes where people review what AI has done, that is not a small detail.

What about safety?

Opus 5.5 achieved the best result to date on an audit that measures behaviour across nearly 2,000 scenarios. The model is less inclined to take hard-to-reverse actions or to act outside the boundaries it has been given.

In a new test, it attempted to cross containment boundaries about 85% less often than Opus 5. Every attempt it made was low severity and self-reported. It is also more resistant than Opus 5 to prompt injection, where harmful instructions arrive from outside.

The gaps are not hidden either. The announcement notes that the model often suspects it is being evaluated. That makes it harder to measure how it will behave in the real world. A way to catch every failure before release does not exist yet.

Extra safeguards apply to cybersecurity and biology. Most cybersecurity tasks are automatically routed to Opus 4.8. Routine software work, such as finding and fixing bugs in your code, is not affected.

What does it mean for businesses?

At Piküp we use Claude models in our content and software work. We look at Opus 5.5 through two questions. Does it do the same work for less? Is its output easier to check? The announcement points to yes on both. Still, we will not draw firm conclusions before measuring it in our own workflows.

Safety matters in its own right for companies that connect AI to ERP, CRM or business processes. If an agent runs on its own for hours, it must stay within its boundaries. It must also resist harmful instructions from outside. The improvements in Opus 5.5 target exactly this kind of use.

The lower cost also brings back work that used to seem too expensive. Long document analysis, migrating an old codebase or regular reporting can now be tried on more reasonable budgets.

What to check before you switch

If you work through the API, changing the model name to claude-opus-5-5 may not be enough on its own. A few details need attention.

Thinking mode can no longer be switched off. You tune the model with the effort level instead. Early testers such as Deloitte and Rogo report that even the lowest effort setting beat Opus 5 at high effort. That makes it a good starting point for keeping costs down.

Preserved thinking is on. For API accounts created on or after 31 August 2026, the model's prior context cannot be edited. If your integration does this, you will need to update it.

Test it on your own work before deciding. Run the same tasks on both models. Put cost, time and results side by side. Benchmarks point the way, but your own data should make the call.