Anthropic released Claude Opus 5.5 on September 22, 2026, lowering the price of its high-end Opus model. The commercial question extends beyond benchmark rankings: whether more demanding AI work can now justify its cost.
Anthropic’s Claude Platform pricing documentation, checked on September 27, 2026, lists the following standard API rates. All amounts are in US dollars per million tokens, before additional tool charges or optional pricing modifiers.
Usage category
Opus 5
Opus 5.5
Reduction
Uncached input
$5
$4
20%
Output
$25
$20
20%
Five-minute cache writes
$6.25
$5
20%
Cache reads
$0.50
$0.20
60%
These are model-usage charges, not monthly subscription prices. The largest percentage reduction applies to reading cached material, rather than generating new output.
Anthropic’s prompt-caching documentation explains how applications can reuse previously processed instructions, documents and tool definitions when the cached prompt segments match exactly. An application repeatedly working from the same client brief or codebase context can therefore benefit from cheaper cache reads. Newly added material and generated answers still incur their respective charges.
The comparison with Anthropic’s more expensive Fable 5.1 is larger. Its model documentation lists Fable at $10 for input and $50 for output, making Opus 5.5’s base rates 60% lower for those categories. That comparison alone does not establish equivalent performance on every task.
Separately, Anthropic estimates that Opus 5.5 costs about 40% less to run than Opus 5 on typical workloads at default settings. That estimate combines lower rates with changes in token consumption. It is a workload-level claim, not a uniform 40% discount on every customer’s bill.
Stronger results do not guarantee a smaller bill
External testing supports the capability advance while qualifying the savings argument. In its September 22 evaluation, Artificial Analysis reported an Intelligence Index score of 58 at maximum effort, the highest it had measured at publication. Its evaluations used Anthropic’s default fallback setting.
However, Artificial Analysis also found that Opus 5.5 used approximately 119,000 output tokens per Intelligence Index task at maximum effort, compared with roughly 73,000 for Opus 5 at maximum effort. Despite producing about 1.6 times as many output tokens, its cost per task was approximately level with Opus 5, rather than 40% lower.
The comparisons describe different operating conditions. Anthropic’s estimate concerns typical workloads at default settings. Its technical documentation specifies that Opus 5.5 defaults to medium effort, whereas Opus 5 defaulted to high, and warns that the newer model can think more per turn at the same effort setting.
Early app-building tests show a route to savings
App-building platform Lovable published a more specific efficiency result on September 22. In its internal evaluation, Opus 5.5 took 34% fewer steps and processed 27% fewer input tokens on iterative code-fixing tasks at medium effort. Lovable reported level quality scores between the models for that category.
Lovable compared equivalent low, medium and high reasoning settings using the same code and judges, with each task run at least three times. Those results describe Lovable’s testing environment, not a guaranteed improvement across all software projects.
The potential business effect follows from two separate changes: paying less for each unit of model usage and needing fewer units to finish acceptable work. For a software company, that combination could support more included usage within an existing product price, or reduce the cost of serving the same customers.
For an agency delivering a fixed-price research or development project, lower model costs could create room for additional checking and revision without changing the client fee. The effect on total delivery cost would be smaller, however, if human review and client coordination account for most of the expense.
For businesses considering internal automation, the opportunity is similarly conditional. A workflow that previously cost more to run than the value it produced could become viable at a lower cost per successful result. Neither the launch announcement nor these early evaluations establishes how widely that threshold has already been crossed.
Deployment costs remain part of the calculation
Anthropic’s model documentation lists Opus 5.5 as available through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Availability gives businesses several deployment routes, but existing integrations are not guaranteed to work unchanged.
Anthropic’s technical documentation identifies breaking changes, including errors for requests that disable thinking or force a specific tool call. Applications using those features may need updates before they can benefit from the new rates. Lower inference prices do not remove that implementation work.
Anthropic’s cost-optimisation guidance also distinguishes cost per attempted task from cost per completed task. Failed attempts can still consume tokens, followed by further spending on retries and whatever the failure costs downstream.
The business case therefore rests on accepted results, including the cost of review and rework. Opus 5.5 makes high-end model access cheaper; whether it makes an entire service or operation cheaper depends on what happens between the first request and a usable outcome.
Get new small business insights by email
Practical ideas and useful articles to help you make better business decisions.
Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us
