Claude Opus 5.5 Is Out: What Changed and What It Means for Your Business

Anthropic released Claude Opus 5.5 on 22 September 2026. It is now the model Anthropic recommends starting with for most work. On Anthropic's own benchmarks it scores at or above Claude Fable 5.1, the company's larger flagship, on coding, knowledge work and computer use, while costing $4 per million input tokens and $20 per million output tokens. That is 20% cheaper per token than Opus 5, and Anthropic says it costs around 40% less to run on typical workloads because it also uses fewer tokens to finish the same job.

If you run a business and use AI through ChatGPT, Claude or an agency like us, the short version is this: the same quality of work got cheaper and quicker to produce, again, two months after the last time.

Every figure in this post comes from Anthropic's launch page, Anthropic's developer documentation or the press coverage linked in the sources at the bottom. Benchmark scores are as reported by Anthropic and have not been independently reproduced yet.

The key facts at a glance

A million tokens is roughly 555,000 words, according to Anthropic's documentation. In practice that means you can hand it a full year of customer emails, a long contract bundle or an entire website's worth of copy in one go.

What actually changed from Opus 5

Opus 5 came out on 24 July 2026. Opus 5.5 improves on it in four ways that matter outside a lab.

It is cheaper. Input drops from $5 to $4 per million tokens and output from $25 to $20. Cached input, which is what you pay when the same instructions are sent repeatedly (a chatbot's system prompt, for example), drops from $0.50 to $0.20 per million.

It is faster. Anthropic says Opus 5.5 produces output more than 30% faster than Opus 5. Several outlets, including 9to5Mac and MarkTechPost, repeat the same figure from the launch.

It says less to get the same thing done. A theme running through the launch is shorter, clearer answers. TechCrunch reports that Opus 5.5 cuts back on jargon and puts the important information at the start of a message. Box told Anthropic its answers were 40% less wordy without losing accuracy. Anyone who has asked an AI a simple question and received six paragraphs will notice this.

It is better at checking its own work. Anthropic highlights self-verification and handing parts of a job to sub-agents. For business use this is the change with the biggest knock-on effect, because the expensive part of AI is not the tokens, it is a human catching the mistakes.

How it scores against Fable 5.1 and GPT-6 Astra

These are the headline benchmarks from Anthropic's launch page. Higher is better on all of them. GPT-6 Astra is OpenAI's current flagship, released on 3 September 2026.

Two things stand out. Opus 5.5 beats Anthropic's own bigger model on every row here, which is unusual. And it does not win everything: GPT-6 Astra is ahead on AutomationBench and on the science benchmark. We compare all of these, plus a very different model called Jev, in Opus 5.5 vs Fable 5.1 vs GPT-6 Astra vs Jev.

A benchmark is a test the vendor chose to publish. Treat these as a guide to direction, not a promise about your own use.

Safety testing and the "slowdown" context

Opus 5.5 is Anthropic's first model since chief executive Dario Amodei publicly argued for pacing the frontier of AI development. TechCrunch quotes him saying that "fully addressing the risks requires even more prudence."

According to Anthropic, the model was tested before release by outside evaluators including METR and Frontier Design, and put through an automated behavioural audit of around 2,000 scenarios. Anthropic reports it tied Fable 5.1 for the lowest prompt injection success rate on the Gray Swan benchmark. Prompt injection is when hidden text in an email or web page tries to trick an AI into doing something it should not, so this matters for any AI that reads customer messages.

Pricing in plain English

AI models are priced per token. A token is a chunk of a word, roughly three quarters of one on average. You pay for what goes in (your question plus any documents) and what comes out (the answer).

Prices are in US dollars, as Anthropic publishes them. The Batch API halves them for work that does not need an instant reply, such as overnight report writing.

For a sense of scale, take a website chat conversation that uses 2,000 tokens going in and 1,000 coming out. At Opus 5.5 prices that is just under 3 US cents, so a thousand conversations comes to about $28. The same thousand on Opus 5 would be about $35. For most small businesses the model bill is not the cost that matters. Setting it up properly is.

If you use Claude through a claude.ai subscription rather than the API, Anthropic says Pro, Max, Team and Enterprise plans get higher five-hour usage limits with this release.

What this means for a small business

We build websites, run marketing and set up AI agents and automations for UK businesses, and we use Claude every day to do it. Here is how we read this release for the owners we work with.

The "best model" is no longer the most expensive one. Until now, the rule of thumb was that the top-tier model was worth paying extra for on hard jobs. Opus 5.5 at medium effort now does most of what Fable 5.1 does at under half the per-token price. Anthropic's own documentation now says to start with Opus 5.5 and only move to Fable 5.1 if it falls short.

Shorter answers are a real improvement for customer-facing AI. If you have a chatbot answering enquiries on your website, a model that gets to the point is a better experience for your customer. It also costs less, because output tokens are the expensive ones.

Better self-checking means less babysitting. The practical limit on AI automation in a small business is how often someone has to check it. A model that catches its own errors more often shifts more work into "set it up and let it run", for things like drafting follow-up emails, summarising calls or keeping a CRM tidy.

Nothing you have already built breaks. Opus 5 is still available as a legacy model. If an agency or developer has built something for you on Opus 5, moving it to Opus 5.5 is usually a one-line change followed by some testing, and the bill should go down.

If you are wondering where AI fits in your business at all, start with the jobs that eat your week. Our AI agents and AI automations pages show the kind of work that pays back fastest.

What is coming next

Anthropic describes Opus 5.5 as the first release in the 5.5 generation. Claude Sonnet 5.5 and Claude Haiku 5.5 are due in the coming weeks, according to the launch page. Sonnet and Haiku are the cheaper, quicker tiers, so if the pattern holds, they are the ones that will make everyday AI jobs cheaper still.

Sources

• Introducing Claude Opus 5.5, Anthropic
• Models overview, Claude Platform documentation
• Anthropic releases Opus 5.5 with lower prices and Fable-level performance, TechCrunch
• Anthropic upgrades Claude with new Opus 5.5 model, 9to5Mac
• Anthropic Releases Claude Opus 5.5, MarkTechPost

Last checked against these sources on 22 September 2026.