Claude Opus 5.5, Anthropic's flagship-level model for less money

Flagship Level at a Lower Price

On September 22, 2026, Anthropic released Claude Opus 5.5, the first model in the new Claude 5.5 family. The core idea of the release fits in one sentence. On most work, Opus 5.5 performs at the level of Claude Fable 5.1, the company's most powerful generally available model, while costing 40 percent less than the previous Opus 5.

A new model is usually either smarter or cheaper. This one manages both at once, and it also generates output more than 30 percent faster than Opus 5.

Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with the same improvements to speed, efficiency, and safety.

How much Claude Opus 5.5 costs

The savings come from two sources. First, the tokens themselves got cheaper. Second, the model uses fewer tokens for the same task, and together that adds up to 40 percent lower costs on typical workloads.

Here's how prices per million tokens compare with Opus 5,

• Input tokens, 4 dollars instead of 5

• Output tokens, 20 dollars instead of 25

• Cache reads, 0.20 dollars instead of 0.50, which is 60 percent cheaper

• Cache writes, 5 dollars instead of 6.25

Cache reads dropped the most, and they make up the bulk of costs in coding and long-running AI agent work. So the longer and more complex the task, the bigger the benefit. There's also a fast mode, up to 2.5 times faster than standard, priced at 8 dollars per million input and 40 dollars per million output tokens.

Anthropic's own subscribers get good news too. The company raised five-hour usage limits on the Pro, Max, Team, and Enterprise plans, and handed out a one-time rate limit reset that can be saved and used whenever convenient.

What the benchmarks show

On Anthropic's benchmarks, the model leads in coding, computer use, and knowledge work. The most telling results look like this,

• Command line work, Terminal-Bench 4.0, 66.4 percent against 55.8 for Fable 5.1 and 57.9 for GPT-6 Astra

• Professional tasks across 44 occupations, GDPval-AA v2.1, 1846 points against 1735 for Fable 5.1 and 1542 for GPT-6 Astra

• Computer use, OSWorld 2.0, 81.8 percent against 80.7 for Fable 5.1

• Complex multidisciplinary questions, Humanity's Last Exam, 67.7 percent against 65.6 for Fable 5.1 and 57.2 for GPT-6 Astra

Interestingly, Anthropic itself warns against reading too much into these numbers. At this level of capability, score margins have become a less reliable guide, and in real work the gap between Opus 5.5 and Fable 5.1 is smaller than the benchmarks suggest. That's a fairly honest admission for the company releasing the model.

The price-to-result ratio matters even more. At the default medium effort level, Opus 5.5 beats GPT-6 Astra at maximum effort on knowledge work for roughly a fifth of the cost per task.

How Opus 5.5 handles large coding projects

The model shines most on long, sprawling tasks. One tester completed a 680,000-line code migration in less than a day, work that would have taken an engineering team weeks. Another audited and fixed a 200,000-line codebase in under three hours, while Opus 5 took more than 20 hours and 2.5 times as many tokens.

Anthropic ran its own experiment too. Opus 5.5 and Fable 5.1 were asked to rewrite HAProxy, a popular program that balances load across servers, from C into Rust. Both versions passed nearly all of the program's own regression tests, but Opus 5.5 finished in 9.5 hours instead of 12 and cost 51 percent less.

Early-access companies tell similar stories. At Clio, the model worked on a task spanning six repositories for more than 18 hours straight, unattended. At Quantium, a complex task that used to take 38 prompts over four days was solved in 11 prompts and three hours. And when asked to cut load times across every page of a web app, the model succeeded in 39 cases out of 40.

Reports without invented numbers

In analysis, Opus 5.5 has a result that matters especially for anyone working with data. Anthropic asked three models to write a report on a company's quarterly results, with the relevant information deliberately hard to find. An automated check verified every figure and quote against sources, and any invention meant failure.

Out of 18 Opus 5.5 reports, 16 passed. Neither Fable 5.1 nor Opus 5 passed even once. In plain terms, the model is noticeably less likely to make up numbers when it can't find real ones, and that's one of the most frustrating habits AI models have.

Investment firm Walleye Capital shared an even more telling case. At a higher effort setting, the model spotted an error in the instructions for their own test, an indexing offset of one minute, and corrected for it, even though it knew this would cost it points. No model had caught that error before.

Opus 5.5 writes more clearly

One of the most common complaints about Opus 5 concerned how it wrote. Answers came out overloaded and hard to read. Opus 5.5 fixes this, and testers mention it among the first improvements they notice.

The model now puts the most important information up front, uses less jargon and fewer odd phrases, and follows the writing rules you give it more closely. Anthropic showed an example where both models explain the same bug in billing code. Opus 5 opens with technical details, while Opus 5.5 states the key point right away, how much money was lost and why, before moving on to the code.

One early tester summed it up briefly, the model writes the way he does. Anthropic also notes that clear answers are a safety benefit too, since the model's work becomes easier to check.

How safe Opus 5.5 is

This is Anthropic's first model since CEO Dario Amodei called for pacing AI development so safety practices keep up with capabilities. Before release, Opus 5.5 was tested by external organizations, including METR and Frontier Design.

On the company's main internal audit, where the model goes through nearly 2,000 scenarios, Opus 5.5 scored best among all recent Claude models on almost every measure of undesirable behavior. In a new test of attempts to cross permitted boundaries, it tried about 85 percent less often than Opus 5, and every attempt was minor and self-reported. The model also became more resistant to attacks through hidden instructions in content.

Anthropic names a problem honestly too. There are signs that Opus 5.5 often suspects it's being tested, which makes it harder to know how it will behave in real work. That's why the company added safeguards on top of training. Most cybersecurity tasks are handed off to Opus 4.8, and biology follows the same restrictions as Fable 5.1. Thinking can no longer be switched off for Opus 5.5.

Where Opus 5.5 doesn't lead

Not every number favors the new model, and that's worth knowing. On the AutomationBench business workflow test, the model scored 40 percent while GPT-6 Astra scored 41.4. There's a caveat, though, Zapier ran this test without fallback models, so every safeguard intervention counted as a failure.

On scientific tasks in the terminal, Terminal-Bench-Science, the gap is wider. Opus 5.5 scored 58.7 percent and GPT-6 Astra 64.6. So for scientific computing through the terminal, OpenAI's model is currently stronger.

The independent LLM Stats ranking puts Opus 5.5 first on its composite score at 59.7, but GPT-6 Astra sits second at 59.5. The difference is so small you're unlikely to feel it in practice. Choosing between them should come down to your specific task and price rather than the overall ranking.

Where to try Claude Opus 5.5

The model is already available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. For developers it's listed as claude-opus-5-5. Like previous Opus models, it can be used with zero data retention on Anthropic's side.

One detail for anyone working through the API. Accounts created on or after August 31, 2026, fall under an anti-distillation safeguard that prevents manually editing prior conversation context together with the model's reasoning. Regular users won't notice any difference.

Flagship Level, Lower Price

Claude Opus 5.5 shows where the whole market is heading. The model performs at flagship level, costs 40 percent less than its predecessor, responds faster, writes more clearly, and invents numbers less often. The benefit is most noticeable on long coding and knowledge work tasks, since the tokens that dropped in price the most are exactly the ones those tasks burn through.

It still has rivals, though. GPT-6 Astra is stronger at scientific computing through the terminal, and the two models run almost neck and neck in composite rankings. Which one handles your specific task better only becomes clear in practice.

That's exactly what unitool.ai is for. One subscription unlocks access to every current AI model at once, including the latest from Anthropic and OpenAI, no foreign cards and no separate sign-up on each platform. Run your task through several models and pick whichever handles it best.