Claude Opus 5.5 Is Out: Faster, Cheaper, Better Scores

Anthropic released Claude Opus 5.5 on September 22, 2026, and it cut prices instead of raising them. Anthropic lists the new model at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5 (Anthropic). The company also says it matches or beats its most expensive model, Claude Fable 5.1, on most of the benchmarks it published.

We run our own work on Claude, so this release affects us directly. Here is what was announced, how to read the numbers, and what a business or a small marketing team should do with it.

What Anthropic announced

• Price. $4 input and $20 output per million tokens. Cache reads fall from $0.50 to $0.20 per million, and cache writes from $6.25 to $5 (Anthropic).
• Total cost. Anthropic estimates Opus 5.5 costs about 40% less than Opus 5 on typical workloads. The list price is only 20% lower. The rest comes from the model using fewer tokens to finish the same task (VentureBeat).
• Speed. Anthropic says it generates output more than 30% faster than Opus 5.
• Where to get it. Anthropic's own platform, AWS, Google Cloud and Microsoft Azure, under the model ID claude-opus-5-5.
• Size limits. A 1-million-token context window and up to 128,000 output tokens per request, according to VentureBeat.
• What's next. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks.

The benchmark results

Anthropic published a comparison table with Opus 5 and Fable 5.1. Selected results, all as reported by Anthropic:

• Terminal-Bench 4.0 (agentic coding in a terminal): 66.4% for Opus 5.5, 55.8% for Fable 5.1, 52.3% for Opus 5 (Anthropic).
• CursorBench 4.0: 57.8% vs 51.8% vs 46.6% (Anthropic).
• GDPval-AA v2.1 (office knowledge work): 1846 Elo vs 1735 vs 1708 (Anthropic).
• OSWorld 2.0 (operating a computer): 81.8% vs 80.7% vs 74.0% (Anthropic).
• Humanity's Last Exam: 67.7% vs 65.6% vs 63.6% (Anthropic).

Two caveats. First, the company selling the model picked the benchmarks and ran them, which is normal for a launch but means the table is a claim until independent testers repeat it. Second, the gaps vary. On computer use, Opus 5.5 is about one point ahead of Fable 5.1. On terminal coding, it is more than ten points ahead. The results on coding and agent work are the strongest part of this release. The rest are close.

Why the lower price matters more than the scores

Most small teams don't choose a model based on which one wins a leaderboard. They choose the best model they can afford to run every day. That's why this release matters. Until today, getting Fable-level results meant paying Fable prices. VentureBeat's comparison puts Opus 5.5's list price about 60% below Fable 5.1's (VentureBeat).

Cheaper cache reads matter for anything that reuses the same long instructions again and again. A support bot, a lead-qualification flow and a content workflow built around a fixed brand guide all send the same prompt prefix on every call. That repeated part is where the 60% cut in cache-read pricing applies (Anthropic).

What Anthropic said about safety

This is Anthropic's first model since it publicly called for the industry to slow its pace at the frontier, as Yahoo Finance reported. The company says outside evaluators, including METR, tested the model before release. It also says Opus 5.5 scored best of any model it has run through its automated behavioral audit, and that it tried to get around its operating limits about 85% less often than earlier models. Those are Anthropic's own findings. The external reports are the ones to watch.

What to do this week

1. Rerun one real task, not a demo. Take a job you already send to a model, such as a weekly report, a batch of product descriptions or a code change, and run it on both models. Compare the output and the token count in the usage log. Anthropic's 40% figure is an average (VentureBeat), and your workload may differ.
2. Turn on prompt caching if you haven't. Cache reads are now much cheaper, so a long system prompt costs very little after the first call.
3. Don't rebuild anything yet. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks. If a cheaper model handles most of your volume today, wait and compare against its new version before you move everything to Opus.
4. Keep your checks in place. A better model still makes mistakes that look confident. Any output going to a customer, whether an email, an ad or a price, still needs a person to review it.

The biggest change isn't the scores. Anthropic's top-level results are now priced for everyday use.

Want to know whether AI search tools like ChatGPT, Perplexity and Google's AI Overviews can find and cite your business? Our free website audit checks that. If you'd rather ask a person, message us on WhatsApp.

← Back to all articles