Anthropic's new top model costs 40% less to run than Opus 5 and returns output more than 30% faster. It lands on Pro, Max, Team and seat-based Enterprise plans.
For a rep, the price drop is the part that changes behaviour, because it changes what you can afford to do twice. The work that used to feel expensive to redo — re-running a full account brief after a champion changes, re-scoring a buying committee after a reorg, rebuilding a deal review when the numbers move — now costs about a fifth less per run, and comes back faster. The 40.0% AutomationBench figure is the only number Anthropic publishes that points at multi-step business work rather than code, and even that is a general automation test, not a sales one. So treat the model as a better engine for the long, messy, context-heavy jobs you already hand to Claude, and do not expect it to improve a one-line email rewrite you were already happy with.
A few terms in the announcement do real work. "Cache reads" are what you pay when the model re-reads context you have already sent it — the same account dossier, the same methodology notes — and Anthropic cut that from $0.50 to $0.20 per million tokens, which is the steepest cut in the whole price table. That is the line that matters if you reuse a standing brief across a dozen prompts. "AutomationBench" is a test of whether a model can complete multi-step business tasks end to end without a human nudging it back on track. "GDPval-AA" is scored in Elo, the chess rating system — 1846 for Opus 5.5 against 1708 for Opus 5 — which only tells you it wins head-to-head comparisons more often, not by how much on any particular job.
The short version: the same Claude work now costs less and comes back quicker, and the biggest savings go to people who keep re-feeding the model the same background.
Anthropic released Claude Opus 5.5 on September 22, 2026. The company says it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5," and that it "also generates output more than 30% faster than Opus 5." The gap Anthropic points to on business work is AutomationBench, where it reports 40.0% for Opus 5.5 against 26.9% for Opus 5.
Announced: September 22, 2026.
Availability: Anthropic says it is "increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans." The model runs in Claude Code, Claude Cowork and the Claude Platform API, and through Amazon Web Services, Google Cloud and Microsoft Azure. The API model name is claude-opus-5-5.
Price: $4 per million input tokens and $20 per million output tokens, against $5 and $25 on Opus 5. Cache reads are $0.20 per million, against $0.50; cache writes are $5, against $6.25.
Regions: Not stated.
What it doesn't do: This is a model swap, not a sales product. Nothing here connects to a CRM, drafts a sequence or joins a call, and no seller-specific feature ships with it. Anthropic also undercuts its own scoreboard: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences." None of the benchmarks cited measure sales work.
SDR — Almost nothing changes for you today. Your work is short prompts on thin context, and that is the one place a stronger model earns the least. If you are on Pro, the raised five-hour limit is the real gift: you can run more research passes in a block before you hit a wall.
AE — This is your upgrade. Take the longest prompt you own — the one that eats a discovery transcript, a stakeholder map and last quarter's notes and returns a deal plan — and run it once on Opus 5.5 against your saved output from Opus 5. Judge it on how much you had to rewrite, not on whether it reads better. If the rewrite burden drops, switch and keep the prompt.
Manager — The faster output changes your Monday. Deal reviews that came back after your pipeline call now come back during it. Before you tell the team to switch, run your own inspection prompt on one live deal and check whether the model's risk calls match yours; a model that reasons differently will flag different deals, and you want to know that before it flags them in front of the team.
RevOps — The cache-read cut from $0.50 to $0.20 per million tokens is the line to model. If you run anything that re-sends a standing context block — a methodology definition, a territory dossier, a product matrix — your per-run cost falls furthest there. Re-price your existing Claude workloads against the new table before you approve any new seats, and note that Anthropic publishes no region list, so if you have a data-residency requirement you need to ask rather than assume.
The honest test of a new model is not a benchmark, it is your hardest recurring job. Take a multi-step deal prompt — a buying committee map, a MEDDPICC inspection, a stalled-deal recovery plan — and run the same one on both models with the same inputs, then count the edits.
Two minutes, once a week. What changed in AI, and what to run because of it.