Claude Opus 5.5 is cheaper, but check your integrations first

Claude Opus 5.5 gives businesses a reason to revisit the cost of their AI workflows. It also gives the people maintaining those workflows a reason to check the integration before changing the model.

Anthropic announced the model on 22 September 2026. Its release notes record availability that day through the Claude API and supported AWS, Google Cloud and Microsoft Foundry offerings. The direct API identifier is claude-opus-5-5. Claude release notes

Mimir's recommendation is to evaluate it on a contained workflow first. A lower token price is useful; a working process with predictable costs and review requirements is more useful.

What Claude Opus 5.5 costs

Anthropic lists standard input at US$4 and output at US$20 per million tokens, compared with US$5 and US$25 for Opus 5. Cache reads cost US$0.20 per million tokens. Cache writes have separate charges. These are API rates, not subscription prices. Model details and pricing

The headline claim of 40% lower running costs comes from Anthropic's tests at default settings. Treat it as a vendor-reported workload result, not a guaranteed reduction in your bill. Launch announcement

For a useful comparison, measure one completed, accepted piece of work. Include retries, tool charges and the time somebody spends correcting the answer. Keep input material and acceptance criteria consistent. Record cache usage separately so a favourable test does not depend on reuse that your normal workflow rarely achieves.

A cheaper request that needs another round of checking might still be worthwhile. It simply needs to be judged on the whole task.

Where existing integrations need attention

Anthropic documents several changes for applications using the Messages API:

• Thinking is always enabled. Requests that disable it or specify a manual thinking budget are rejected.
• Forced tool choices using any or a named tool are rejected; the accepted choices are auto and none.
• On the Claude API and Google Cloud, the older computer_20251124 tool must be replaced with computer_toolset_20260801.

These details are platform-specific. Check the migration instructions for your starting model and hosting route. Migration guide

There is also a quieter interface change: progress text between tool calls now arrives inside thinking blocks and is hidden at the default display setting. A workflow can continue running while its interface appears silent. Behaviour changes

For the business owner, this means the upgrade belongs on a short change plan. For the developer, it means testing request settings, response parsing and the user interface together.

Run a pilot that can give you a clear answer

Our suggested pilot uses familiar, sanitised work: summarising an internal document, drafting a support response or proposing a small code change. Keep customer communications and production changes under human approval.

Before the trial, write down what an acceptable result looks like. Then assess the existing model and the candidate against the same requirements:

• Accuracy: Are the important statements supported by the supplied material?
• Completion: Did it finish the task, including awkward inputs and failure cases?
• Effort: How much checking and correction did a person need?
• Operation: Did permissions, tool calls, progress updates and recovery behave as expected?
• Cost: What did an accepted result cost, including repeated attempts?

Keep a known working version of the integration available. Expand the trial only when the team can explain both the improvement and the remaining limitations.

What this briefing does not establish

Mimir has not independently tested Opus 5.5. The linked system-card document could not be retrieved during this scan, so this briefing does not reproduce benchmark scores or draw safety rankings.

If your project involves an agent acting across business systems, our existing guide to AI agent failure modes covers the wider operational questions. For deciding where to begin with repeatable processes, see workflow automation in Microsoft 365.

For an existing Claude integration, the practical next step is a small, reversible evaluation. Approve the change when the resulting work, review time and operating cost justify it.