Opus 5.5 Dropped Today. Here Is What It Actually Changes for Law Firms.

Anthropic released Claude Opus 5.5 today. Input dropped to $4 per million tokens, output to $20, and cache reads fell 60 percent to $0.20. It generates output about 30 percent faster than Opus 5, and it beats Fable 5.1 — the larger, more expensive model — on most benchmarks Anthropic published. Terminal-Bench 4.0: 66.4 percent against Fable’s 55.8. Knowledge work on GDPval-AA: 1,846 Elo against 1,735. Computer use on OSWorld 2.0: 81.8 against 80.7.

Those numbers are real. What they mean for your firm is narrower than the headlines suggest, and it is worth being precise about.

The cheapest model is not the point

Opus 5.5 is not the cheapest Claude model. Sonnet 5 runs $2 in and $10 out. Haiku 4.5 runs $1 and $5. Measured per token, Opus is still four times Haiku. Anyone telling you the new Opus is the cheapest option for daily firm work is reading the press release instead of the price sheet.

The number that matters is cost per completed task. Anthropic reported that across customer deployments, Opus 5.5 needed roughly half the steps Opus 5 needed to finish the same terminal work. One company cut a four-day, 38-prompt job down to 11 prompts over three hours. When a model needs half the turns to reach a correct answer, the per-token discount is the smallest part of the savings.

That distinction is the entire argument for frontier models in legal work. A cheap model that gets a financial affidavit cross-check 80 percent right bills you the other 20 percent in attorney review time, and attorney review time is the most expensive line in the firm. Model selection has never really been about the sticker price. It is about how much of the work comes back to your desk.

Where this actually lands in a law firm

Most attorneys are not buying tokens. They are paying a flat subscription, which means a 20 percent API price cut does not change their bill by a dollar. What changes is what that same subscription can carry.

Long-horizon work is where the gap opens. Discovery review across 400 pages of bank statements. Building an equitable distribution chart from raw production. Cross-checking a financial affidavit against three years of returns and a paystub history. Those tasks fail on weaker models not because the model cannot read a bank statement, but because it loses the thread somewhere around step forty. One early tester handed Opus 5.5 a job spanning six repositories and let it run overnight, unattended. It stayed on task for more than eighteen hours. Eighteen hours of sustained focus is not a coding statistic. It is the reason a discovery pipeline finishes in one pass instead of three.

The writing changed too. Anthropic went after what testers called Claudish output — the padding, the throat-clearing, the restating of your question before answering it. Opus 5.5 leads with the finding. If you have ever cut an AI-drafted letter in half before it was usable, that is a real hour back every week.

What did not change

The model got better. Your obligations did not move an inch. Opus 5.5 will still hand you a citation that does not exist if you ask it to research caselaw without giving it caselaw. Competence and supervision duties read exactly the same today as they did yesterday. Every citation still gets attorney verification. Every filing still gets attorney sign-off. Check your judge’s standing order on AI disclosure before anything reaches the clerk, because a faster model produces noncompliant filings faster.

And a better engine does not fix a system that was never built. Most firms disappointed by AI are not using the wrong model. They are using a general-purpose chatbot with no jurisdiction loaded, no forms, no procedural rules, and no workflow behind it. Opus 5.5 with nothing built around it is still a very fast intern who has never read the Florida Family Law Rules of Procedure.

What to do this week

Nothing urgent. If you are on a paid Claude plan, you already have it — open your model selector and confirm. If you run anything through the API or through a built system, check which model ID that system is pinned to, because a workflow pointed at an older Opus will keep running on an older Opus until someone changes it.

Then re-run one task you gave up on. Pick the one that almost worked — the discovery summary that drifted, the chart that came back half-built, the response set that needed three rounds of correction. That is the honest test of a generational jump, and it takes twenty minutes. Benchmarks are someone else’s cases. Your file is the benchmark that matters.

Legal Authority Lab builds AI workflow systems for attorneys and law firms. Not built by technologists who have never set foot in a courtroom — built inside a real, active Florida family law firm and tested on real cases, real clients, and real deadlines. AI built for law firms. Powered by real lawyers.

Sources: Anthropic, Introducing Claude Opus 5.5; Anthropic Claude Platform pricing documentation; TechCrunch and The Decoder coverage, September 22, 2026. Nothing here is legal advice, and nothing here substitutes for independent attorney verification of any AI output.