Claude Opus 5.5 and Sonnet 5.5: What Changes for Businesses

In the space of a week, Anthropic released two new Claude models: Opus 5.5 on 22 September and Sonnet 5.5 on 28 September 2026. The headline is not a leap in raw intelligence. It is that capable models got noticeably cheaper and faster, which changes the maths on a lot of business automation.

What did Anthropic actually release in September 2026?

September was a busy month. On 1 September Anthropic released Claude Fable 5.1 and Mythos 5.1, updates to its most capable model family. Fable 5.1 is generally available; Mythos 5.1 remains restricted to vetted cyber defenders and life scientists. Fable 5.1 kept its list price of $10 per million input tokens and $50 per million output tokens, but Anthropic says it costs roughly 25% less than Fable 5 on typical workloads.

Three weeks later came Claude Opus 5.5, the first model in the 5.5 family, followed on 28 September by Claude Sonnet 5.5. Anthropic describes Opus 5.5 as performing at the level of Fable 5.1 on most work, at a lower price. Sonnet 5.5 is the everyday model: fast, cheaper and available on the same platforms.

For context, this is the third generation of Claude released this year. Sonnet 5 arrived on 30 June and Opus 5 on 24 July 2026. The pace is relentless, which is itself a lesson we come back to below.

How much cheaper and faster are the new models?

Here are the numbers Anthropic published at launch:

• Opus 5.5: $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. Anthropic estimates typical workloads cost around 40% less than on Opus 5, partly because the model uses tokens more efficiently, and output is generated about 30% faster.
• Sonnet 5.5: $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. Anthropic says it generates output over 30% faster and costs up to 30% less per task, because it gets to the answer with less work.

Treat vendor benchmarks with healthy scepticism, but the direction is clear. The price of a given level of capability keeps falling, and the fall is steepest for the kind of multi-step, tool-using work that business automation depends on.

What are these models good at, in business terms?

Anthropic positions Opus 5.5 for agentic coding, computer use and knowledge work. One early tester, according to the announcement, completed a 680,000-line code migration in less than a day. On GDPval-AA, a benchmark built around realistic professional tasks, Anthropic reports Opus 5.5 scoring above Fable 5.1.

Sonnet 5.5 is pitched at the daily grind: fixing bugs, drafting documents, building spreadsheets and slides from templates, and customer support style work. Anthropic says it scores nearly level with Opus 5.5 on GDPval-AA. If that holds up in your own testing, the cheaper model becomes the default for a large share of routine tasks.

Translated into the work we see in Maltese and European businesses, that means:

• Reading and extracting data from invoices, contracts, bills of lading and application forms (see processing documents with AI).
• Drafting first versions of reports, proposals and client emails that a person then reviews.
• Running multi-step back-office processes that touch several systems, such as reconciling a supplier statement against your accounting software (see automating entire workflows).
• Helping small development teams maintain and modernise internal software faster.

Are the new models safer to put near real data?

Anthropic says Opus 5.5 matches or beats Opus 5 against prompt injection in every setting it tested, and ties Fable 5.1 for the lowest success rate. Prompt injection is the trick where instructions hidden inside a document or web page try to hijack an AI agent. That matters more than any benchmark if you plan to let a model read supplier emails or browse the web on your behalf. It does not remove the need for sensible design: limited permissions, human approval for anything irreversible, and logging.

Sonnet 5.5 is also, as The Next Web reports, the first Sonnet model to ship with the frontier-style cyber safeguards Anthropic previously reserved for its top models. Separately, Anthropic has published how its text watermark works. With the EU AI Act’s transparency rules on AI-generated content applying from 2 August 2026, provider-side marking is worth understanding, although it does not remove your own disclosure duties as a deployer.

Should you switch your existing workflows to the new models?

Not automatically. Our view is that a model change should be treated like any other software upgrade:

1. Keep a test set. Collect 30 to 100 real examples of the task (anonymised where needed) with the correct answers. Without this you are guessing.
2. Run old and new side by side. Compare accuracy, cost per task and time per task, not just the monthly bill.
3. Try the cheaper tier first. If Sonnet 5.5 matches your current Opus results on your test set, the saving can be substantial.
4. Check where your data is processed. Access through AWS, Google Cloud or Azure may let you keep processing within the cloud region and contracts you already use.
5. Switch gradually. Route a share of traffic to the new model, watch the results, then move the rest.

If your automations were built with the model hard-wired into the code, this is the moment to fix that. Good systems let you swap models with a configuration change, which is part of how we approach connecting systems for clients.

What does this mean for smaller businesses?

The biggest shift is in which projects make financial sense. A task that was marginal at last year’s prices, such as reading every inbound supplier email and updating your stock system, may now pay back comfortably. Faster models also make interactive tools feel responsive enough that staff actually use them.

The other lesson is about pace. Anthropic alone has shipped three generations of Claude in 2026. Any plan that depends on one specific model will date quickly. Build around your processes, your data and your controls, and treat the model as a replaceable component. That is the approach behind our AI automation work.

Frequently asked questions

If you want to know which of your processes the new models make worth automating, book a free discovery call with Haystack. We will look at your workflows and give you a straight answer.