Claude Opus 5.5 launches: access, limits and API prices

Anthropic has released Claude Opus 5.5, with a practical pitch for people already paying for Claude: more room to work before hitting a usage limit, alongside cheaper API processing for developers. The company’s announcement is dated 22 September 2026.

What changes if you use Claude for work?

Anthropic says the new model produces output more than 30% faster than Opus 5 and writes more clearly. Those are company claims. Faster output is not a promise that a research task, spreadsheet or code change will finish 30% sooner.

The useful test is the work you keep having to correct. Take a previous answer that buried the important point, missed an exception or needed several follow-ups. Give the new model the same material and instructions, then compare how much editing remains. A smoother first impression is welcome; less repair work is more valuable.

There is also a rate-limit reset for subscribers, which Anthropic says can be saved for later. If you have an uninterrupted afternoon of work planned, check the reset offer in your account before spending it on a quick experiment.

What the lower API prices actually mean

The Claude API documentation lists standard Opus 5.5 rates of US$4 per million input tokens and US$20 per million output tokens. Input is the material sent to the model; output is what it generates. These are usage charges for software accessing Claude, separate from a consumer subscription.

Against Opus 5’s listed US$5 input and US$25 output rates, each falls 20%. Anthropic’s larger 40% typical-workload saving also reflects efficiency and caching. The percentage your application saves will depend on the work it does.

For a simple illustration, 100,000 input tokens and 10,000 output tokens at the standard Opus 5.5 rates cost US$0.60. That calculation excludes caching, tools and other charges, and is not an estimate of how many tokens your task will need.

Australian readers should keep the currency in view: those figures are US dollars. A lower per-token rate also does not settle the whole cost comparison if you change the task, reasoning settings or number of attempts. Compare completed, checked work on the same basis.

Why a conversation might switch to an older model

Claude’s help documentation describes a safety feature that can move some conversations to a less capable model when a classifier identifies higher-risk work. With Opus 5.5, certain cybersecurity requests can fall back to Opus 4.8; biology and frontier AI-development work can fall back to Opus 5.

Automatic switching is enabled by default in Claude’s apps. The help page says Claude displays a notice and the conversation continues with the fallback model. The API does not enable automatic fallback by default.

That gives you something specific to check if a response suddenly feels different: look for a switching notice and the model name before assuming the newer model has become less capable.

Should you change your workflow?

If Opus 5.5 is already included in your plan, start with a contained task whose answer you can check. For a document summary, pick one with a date change or an exception in the small print. Ask for those details explicitly and compare them with the original. This is a suggested evaluation, not a test we have performed.

Developers should read the migration notes before swapping in claude-opus-5-5. Anthropic documents breaking changes, including always-on thinking and errors for forced tool use. A model that answers a test prompt successfully can still break an existing integration.

Related: what we previously confirmed about the Opus release rumours

For now, the strongest reason to try the release is straightforward: see whether it handles a familiar job with fewer corrections. Keep the original files and check the answer before putting it to work.