Anthropic released Claude Opus 5.5 on 22 September 2026, two months after Opus 5. The pitch fits in one line: results at the level of Claude Fable 5.1, the company’s largest model, at a price 20% below Opus 5 per token and, by Anthropic’s estimate, 40% below it on typical workloads. The model is live in the Claude apps, on the Claude Platform as claude-opus-5-5, and on Amazon Web Services, Google Cloud and Microsoft Azure.
TechCrunch, 9to5Mac and Yahoo Finance all led with the same thing: this is not only a stronger model, it is a cheaper one. That combination is what makes the release relevant to businesses that are not AI labs.
What Claude Opus 5.5 costs
Anthropic’s launch page lists the API prices per million tokens:
- Input: $4 (Opus 5: $5).
- Output: $20 (Opus 5: $25).
- Cache reads: $0.20 (Opus 5: $0.50), 60% lower. Cache writes cost $5.
- Fast mode: $8 input and $40 output, up to 2.5 times faster.
The bigger saving, according to Anthropic, comes from the model doing the same job in fewer steps and fewer tokens. Output is 30% faster than Opus 5. Customer quotes on the launch page put numbers on it: Optiver says Opus 5.5 matched Opus 5’s quality “in about half the turns, time and output tokens”, and Kiro reports about 40% fewer calls using half the tokens.
For people on subscriptions, 9to5Mac notes that five-hour usage limits went up on the Pro, Max and Team plans, and subscribers now get a rate limit reset they can save and use when they choose.
How it compares with Fable 5.1, Opus 5 and GPT-6 Astra
Anthropic published a comparison table. These are the vendor’s own measurements, not independent tests, so treat them as a starting point:
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1846 | 1735 | 1708 | 1542 |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% |
| OSWorld 2.0 (computer use) | 81.8% | 80.7% | 74.0% | — |
| Humanity’s Last Exam (with tools) | 67.7% | 65.6% | 63.6% | 57.2% |
Two details are worth noting. GPT-6 Astra still leads on AutomationBench and on Terminal-Bench-Science (64.6% against 58.7%), so Opus 5.5 is not ahead everywhere. And 9to5Mac points out that on FrontierCode Opus 5.5 edges out Astra at roughly 20% of the cost per task, which matters more than the score itself.
What Anthropic says changed in practice
Long jobs. Anthropic cites a 680,000-line code migration finished in less than a day, and a test where the model was asked to cut load times across every page of a web app: Opus 5.5 succeeded 39 times out of 40, while Opus 5 made smaller improvements that also changed how the app behaved.
Clearer writing. The company says the model puts the important information first and uses less jargon. Box reports output 40% less verbose than Opus 5 without losing accuracy. TechCrunch highlighted the same change.
Analysis and checking. Deloitte says Opus 5.5 caught 72% of known bugs against 56% for Opus 5 at high effort, with fewer false alarms. Thomson Reuters Labs and Hebbia both report 86.6% coverage against 60.3% for Opus 5 on their research tasks.
Safety. Anthropic reports the best scores of any recent Claude model on its automated behavioural audit, 85% fewer attempts to get around containment boundaries than Opus 5, and better resistance to prompt injection. The model was evaluated before release by METR and Frontier Design. Yahoo Finance notes it is Anthropic’s first model since CEO Dario Amodei published an essay this month calling for a slowdown in AI development.
What comes next. Anthropic calls this the first release of the 5.5 wave: Sonnet 5.5 and Haiku 5.5 are due in the coming weeks.
What it means for your business
The price of “good enough” dropped again. A month ago the choice was a frontier model at a frontier price or a cheaper one with more rework. If Anthropic’s figures hold on your tasks, Fable-level quality now costs Opus money. Our view: this is the moment to rerun the three tasks you actually care about and compare cost per finished task, not price per token. AI vendors are competing on the cost of getting work done, from models to ChatGPT’s sponsored agents, and your comparison should be measured the same way.
Automations that were too expensive may now pay off. Lead routing, enquiry triage, CRM hygiene, weekly reports: these run thousands of times a month, and a 40% lower bill changes the business case. Most of our CRM and marketing automation projects stall on exactly this calculation, so it is worth redoing it with the new prices.
Speed matters most on websites and code. A model that finishes a site-wide performance task in one pass instead of several is useful to anyone running a web store or a multilingual site. For our own website development work the practical gain is fewer review rounds, not magic.
Do not switch on a headline. Vendor benchmarks are measured on the vendor’s setup. Test on your own data, keep the old model as a fallback for a week, and watch where the new one makes a different call. That is especially true for customer-facing text, where a shorter, more direct style is an improvement only if it still sounds like your brand.
A note on transparency: M-ADS uses Claude in its day-to-day work, and this article was researched and drafted with Claude Opus 5.5 on launch day, then checked against the sources listed below. We have no commercial relationship with Anthropic.
