Claude Opus 5.5 Helped Write This. The Line I Care About Is 16 Out of 18.

This article was prepared with Claude Opus 5.5. So was the batch of posts it belongs to: the research, the first drafts, the rewrites after I changed my mind halfway through. Anthropic released the model on September 22, and I have used it since in Claude Code, on the kind of long, messy sessions where a model either holds the thread or quietly drops it.

I want to talk about one line in the launch post, because it is the line that decides how I use it.

Sixteen out of eighteen

Below the benchmark tables, Anthropic describes an internal test. Opus 5.5, Fable 5.1 and Opus 5 were each asked to write a report on a company's quarterly performance, using only what they could find on a copy of the web where the earnings release was hard to locate. An automated grader checked every figure and every quote against the sources. One invented number or one invented quote, and the report failed.

Across different effort settings, 16 of Opus 5.5's 18 reports cleared the bar. Neither Fable 5.1 nor Opus 5 cleared it in any attempt.

That is a real step. Going from zero clean reports to sixteen is not a benchmark margin, it is a change in behaviour: when the source is hard to find, the new model is far more likely to keep looking, or to say it did not find it, than to fill the gap with something plausible.

Now read it the other way. Two reports out of eighteen still contained at least one invented figure or quote. One in nine.

One in nine is a rate, not a verdict

For a chatbot answering casual questions, one in nine might be tolerable. For an article with my name on it, or for the legal and IP work the product I run as CTO does every day, it is not. The difference is not the model. It is the cost of the error, which is carried by whoever signs.

So nothing changes in my process, and I suspect Anthropic would agree, since the test only exists because they did not assume. The model writes. Something else checks. In my case, that something else is a small judge model that reads each sentence against the sources, and plain code that recomputes every number. I will write about that setup separately. The point here is that a better writer does not remove the need for a reader. It just makes the reader's job quieter.

What changed in daily use

Two other claims in the launch post match what I have seen.

The first is writing. Anthropic says Opus 5.5 communicates more naturally, puts the most important information up front, is less likely to use jargon and "follows the writing rules you give it." My rules are specific and slightly obsessive: no em dashes anywhere, primary sources only, dates written out, every figure traceable. Halfway through this batch I added a new one: stop citing a particular newspaper, even as context. It applied the rule to everything that followed without being reminded. That sounds small. On a batch of twelve pieces, it is the difference between reviewing and rewriting.

The second is cost, and specifically the cache. Cache reads fall from $0.50 to $0.20 per million tokens, 60 percent less than Opus 5, and Anthropic notes that cache reads make up the majority of the cost of agentic and coding work. A long session is mostly the model rereading the same context: the rules, the files, the notes. That is exactly the line that got cheapest. Input and output tokens are 20 percent cheaper, at $4 and $20, and Anthropic says the model also uses fewer tokens per task, for a 40 percent drop on typical workloads.

The first release after "pacing the frontier"

There is one more sentence in the announcement worth sitting with. Anthropic calls Opus 5.5 its first release since it called for "pacing the frontier," a reference to Dario Amodei's essay of September 12, which argued that labs should slow the rate at which they improve the capabilities of their models, and committed Anthropic to hosting embedded outside evaluators.

Ten days later, the company shipped a model. You could read that as a contradiction. I read the launch post differently. Opus 5.5 is presented as performing at the level of Fable 5.1, Anthropic's existing top model, on most work, while costing 40 percent less to run than Opus 5. It was tested before release by external evaluators, including METR. Anthropic says it posts the best scores it has recorded on its automated behavioural audit and is much less likely than recent models to take hard-to-reverse actions. And because it is comparable to Mythos 5.1 in biology and cybersecurity, it ships with safeguards similar to those on Fable 5.1, with some tasks handed to older models when those safeguards step in.

In other words, the headline gain of this release is efficiency, not a new ceiling. Whether that is what pacing looks like in practice is a question for the evaluators Amodei wants inside the building. For a builder, the practical effect is simpler: the capability I was paying top-tier prices for a month ago now costs less and, by Anthropic's measure, generates output more than 30 percent faster than Opus 5.

What I would tell another CTO

Switch, measure, and keep the checker. The model is better at not making things up, it is not perfect at it, and the vendor published the number that tells you both. That is the most useful kind of launch post: one that gives you the rate, so you can decide what it costs you when the rate hits.

Sources

• Anthropic, "Introducing Claude Opus 5.5", September 22, 2026 : the 16 of 18 test, communication claims, prices and cache share of costs, external evaluators, safeguards, first release since pacing the frontier
• Dario Amodei, "We Must Pace the Frontier", September 2026 : the case for pacing the frontier
• CNN, September 12, 2026 : date of the essay