Claude Opus 5.5: Anthropic's Smarter, Cheaper Opus
By Chatday Editorial Team ·
Claude Opus 5.5 is Anthropic’s new top Opus model, and it costs less to run than the one it replaces. It came out on September 22, 2026, and it’s built for long, hard jobs: big coding projects, research reports and work that takes hours rather than seconds. You can chat with it in Chatday today.
What Claude Opus 5.5 is, in plain words
Anthropic sells Claude in a few sizes. Haiku is the quick, light one. Sonnet is the everyday all-rounder. Opus is the heavyweight you bring in for the difficult stuff. Opus 5.5 is the newest heavyweight, and it’s the first release in a new 5.5 generation. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.
The pitch is simple. It does more of the hard work on its own, and it does it for less. Anthropic describes it as a big step up from Opus 5 for coding, for long tasks an AI carries out step by step on your behalf (the industry calls these “agentic” tasks), for knowledge work like analysis and reports, and for understanding images.
You can see what it’s like on its Claude Opus 5.5 model page, which lists its specs next to every other model on the hub.
The price cut is the part most people will feel
New top models usually cost more. This one costs less.
| Claude Opus 5 | Claude Opus 5.5 | |
|---|---|---|
| Price to send text in (per 1M tokens) | $5 | $4 |
| Price to get text out (per 1M tokens) | $25 | $20 |
| Reading from saved memory, or “cache” (per 1M tokens) | $0.50 | $0.20 |
| How much it can read at once | about 1M tokens | about 1M tokens |
A token is a small chunk of a word, so a million tokens is roughly a few long novels. Those prices matter mostly to developers who pay per use. Anthropic’s bigger claim is about the total bill: because Opus 5.5 also finishes tasks in fewer steps, it says typical workloads come out around 40% cheaper than on Opus 5, and that answers arrive more than 30% faster.
What it’s actually better at
Anthropic shared a stack of test results. Here’s what they mean once you strip out the jargon.
Big, boring coding jobs. Anthropic says an early customer moved a 680,000-line codebase to a new setup in under a day, and that a 200,000-line code review took under 3 hours compared with more than 20 on Opus 5. Those are the vendor’s own examples, but they show the direction: less babysitting on long tasks.
Research and reports. On an internal test of written, sourced reports, Anthropic says Opus 5.5 passed 16 of 18, while the earlier models it was compared against passed none.
Clearer writing. This is the change regular users are most likely to notice. Anthropic says it leads with the key point, uses less jargon and actually follows the style rules you give it. Box, an early tester, said its answers came out 40% less wordy without losing accuracy.
Reading screens and charts. It’s better at understanding images, screenshots and charts, and at working through software the way a person would.
Anthropic also says the real-world gap between Opus 5.5 and its most powerful model is smaller than the test scores suggest. That’s an unusually modest note to put in a launch post, and worth keeping in mind before treating any benchmark as gospel.
It always thinks first, and you can’t switch that off
Opus 5.5 uses what Anthropic calls adaptive thinking: before answering, it works through the problem privately, spending more effort on hard questions and less on easy ones. The default is a medium amount of effort, and thinking can’t be turned off.
In practice, that means it’s a little slower to start than a chatty, instant model, but it’s less likely to fire off a confident wrong answer on a tricky question. If you just want a quick reply to a simple question, a lighter model may feel snappier. We compared that trade-off in our guide to fast vs smart AI models.
Safety: what Anthropic changed
This is Anthropic’s first release since its CEO, Dario Amodei, called for pacing the frontier of AI. Outside evaluators, including METR, tested it before launch. Anthropic says it was about 85% less likely than Opus 5 to try to get around the limits placed on it in a dedicated test, and that it resists hidden malicious instructions (so-called prompt injection) better than Opus 5.
There are guardrails, too. Most cybersecurity tasks get routed to an older model, and some biology work is limited to vetted organizations through a verification program. If you’re writing, coding or researching, you won’t run into any of this.
Should you switch to Opus 5.5?
| If you mostly… | Good pick |
|---|---|
| Work on long coding or data projects | Claude Opus 5.5 |
| Write reports, analyses or long documents | Claude Opus 5.5 |
| Want quick everyday answers | A lighter, faster model |
| Are happy with Opus 5 and want to check the difference | Compare them side by side |
The easiest test is a real one. Take a task you’d normally give Opus 5, run it on both, and see which answer you’d actually use.
If you’re curious about the company behind it, we covered how Anthropic became one of the most valuable AI startups.
Give it the job you’ve been putting off: the long report, the messy spreadsheet, the code you don’t want to touch. That’s exactly where Opus 5.5 is meant to earn its keep.