Box, a company that stores business files online, tested Claude Opus 5.5. Its answers were 40% shorter than the older version's answers, and they were just as accurate.

Claude is an AI assistant, like ChatGPT. A US company called Anthropic makes it. Businesses use the most powerful version of Claude, called Opus, for demanding work such as research, reports and financial analysis. Anthropic released Claude Opus 5.5 on 22 September 2026. It is the upgrade to the previous version, Opus 5.

Most news stories about the launch will focus on how the new AI scored in tests. The change you will notice every day is how it writes. Anthropic says the way Opus 5 wrote was one of the most common complaints it heard from users. So Anthropic changed how the new version writes, in three ways:

It puts the answer first. The most important information comes in the opening lines.

It uses less jargon. It is less likely to use technical terms and odd phrases.

It follows your writing rules. If you tell it how you want things written, it sticks to those instructions.

A shorter answer is a cheaper answer.

Four terms, in plain English

AI companies describe their products in technical language. These four terms are all you need to follow the rest of this article.

What changed in the writing

Anthropic published examples of the old model and the new model answering the same question. In one example, a business asked why a customer's August bill was lower than expected. The two models found the same cause. They explained it in a different order.

The old model gave the technical detail first and left the reader to work out the answer. The new model gave the answer first and put the technical detail after it.

Here is why that matters to you. The person reading an AI answer is usually busy, and is often the person paying for it. When the answer comes first, you can read it quickly and act on it. When the answer is buried, you are more likely to skim it, misread it or ask the AI to rewrite it. An engineer at Ramp, a US finance software company, described the new model this way: "It writes like a good colleague."

Shorter answers, just as accurate

You might expect a shorter answer to leave things out. Businesses that tested Opus 5.5 before the launch found the opposite. The new model wrote less and got more right.

Deloitte, the consulting firm, asked both models to check software for known mistakes. The new model found more mistakes on its cheapest setting than the old model found on a high setting. It also raised fewer false alarms and wrote much less.

Rogo, a company that builds AI tools for finance firms, found the same pattern. It ran a finance test. Opus 5.5 on its cheapest setting scored higher than Opus 5 on a high setting. It also wrote about 60% less text.

Chicago Trading Company said the documents the new model wrote needed almost no editing. That matters because editing costs you twice. You pay for the AI to write the answer, then you pay again in your own time to fix it.

Why shorter answers cost less

AI companies charge by the amount of text the AI reads and writes. So clearer, shorter writing lowers the bill. There are three reasons.

Reason one: the price per token has dropped. Anthropic charges US$20 for every million tokens Opus 5.5 writes. Opus 5 cost US$25 for the same amount.

Reason two: the new model writes fewer tokens to do the same job. Box found Opus 5.5 used a third of the tokens Opus 5 used. Anthropic says the lower price and the shorter answers together make Opus 5.5 40% cheaper to run than Opus 5 for typical work.

Reason three: you send fewer requests. Each time you ask the AI to fix or explain an answer, you pay for a new answer. Quantium, the Australian data science company, counted how many requests one large software job took.

You might pay a flat monthly fee for Claude instead of paying per token. In that case, your plan limits how much you can use Claude in each five-hour period. Anthropic has raised that limit on its Pro, Max and Team plans and some Enterprise plans. You can now do more work before you reach it.

One caution. Every figure in this article comes from Anthropic or from businesses Anthropic invited to test the model early. Treat these figures as the maker's claims until independent testers repeat them.

What to do with it this week

Give it your writing rules. Anthropic says Opus 5.5 follows the writing rules you give it. Write down five rules, for example: British spelling, no jargon, the answer in the first line, a maximum length, and a friendly tone. Paste those rules at the start of your request. If you use Claude's Projects feature, paste them into the project's instructions so they apply to every request. Then check the AI keeps to them.

Test it on work you usually rewrite. Find a report summary or client email the AI wrote for you last month that you had to edit heavily. Give Opus 5.5 the same request. Count how many changes you still need to make. That test on your own work tells you more than any claim in a launch announcement.

Start with the cheapest setting. Three testers compared the settings: Deloitte, Rogo and the software company Factory. All three found Opus 5.5 on a low or medium setting did as well as Opus 5 on a high setting, or better. If your AI tool has an effort setting, start on low. Move it up only when an answer falls short. A low setting gives you faster and cheaper answers.

Test it on the work you rewrite most.