Claude Opus 5.5 Costs Less to Run. Should You Switch?

Short answer: Claude Opus 5.5 is cheaper to run than the model before it, but a price drop is not a reason to rebuild anything. Test it on your own tasks first, and switch only where the saving is measurable and the quality holds.

What Happened

Anthropic released Claude Opus 5.5 on 22 September 2026. Reports put the list price at $4 per million input tokens and $20 per million output tokens. Cache reads, which make up a large share of the cost of agent and coding workloads, are reported at $0.20 per million tokens. A faster mode is also reported, at $8 and $40 per million tokens.

Anthropic says typical workloads cost about 40% less to run than on Opus 5. That is a claim about running costs, not a 40% cut in list price. The model also uses fewer tokens to finish a task, and that accounts for part of the saving. Reports differ on what Opus 5 cost at launch, so check Anthropic's pricing page for current figures before you put numbers in a budget.

The model is available through Anthropic's API, through AWS, Google Cloud and Microsoft Azure, and on the Pro, Max, Team and Enterprise plans.

Why the Headline Number Needs Care

A "40% cheaper" headline tells you very little about your own bill. Three things decide what you will pay.

The first is your mix of input and output. Output tokens cost five times as much as input tokens at the reported prices. A workflow that reads long documents and returns short answers behaves very differently from one that writes long reports.

The second is caching. If your system sends the same instructions or reference material again and again, cached reads at the reported rate can change the sums. If it does not use caching, you will not see that benefit.

The third is token use. A model that finishes a task in fewer steps costs less even at the same list price. That only shows up when you measure it on your own work. A benchmark on someone else's tasks will not tell you.

What It Means for an SME

Model cost is rarely the biggest line in an AI project. Integration, data preparation, testing and staff time usually cost more. A cheaper model helps most when you already run high volumes, such as document processing, support triage, quote drafting or internal search.

If your model spend is a few hundred pounds a month, this release changes very little for you. If your team pushes thousands of documents through a system each day, a real saving is worth testing.

If you signed a fixed-price build before this release, ask your supplier whether a price drop like this reaches you or stays with them. A well-built system should let you pass savings through, or at least show you where the costs sit.

What I Would Do This Week

1. Pick one live workflow and write down its current monthly model cost.
2. Collect 50 real examples from that workflow, including a few awkward ones.
3. Run them through the new model and your current one.
4. Compare accuracy first, then cost. A cheaper wrong answer costs more than a dearer right one.
5. Check how your system calls the model. If the model name is buried in many places, fix that before the next release arrives.

Keep the test small. It costs an afternoon and commits you to nothing.

What to Watch For

Do not assume a newer model behaves the same way as the old one. Prompts that worked well can change in tone, length or format. Test the edge cases your staff complain about, not only the easy ones.

Also be wary of switching only because the price fell. Every switch costs testing time, so make it for a measured gain. Switch when the evidence supports it, and build so that switching is easy.

Frequently Asked Questions

Is Claude Opus 5.5 cheaper than Opus 5? Reports say yes. The list price is reported as lower, and Anthropic says typical workloads cost about 40% less because the model uses fewer tokens.

Do I need to rebuild my AI system to use it? Not if it was built with a model-agnostic layer. In a well-built system, changing the model is a configuration change followed by testing.

Should a small business use the most expensive model? Usually not. Match the model to the task. Simple classification and drafting rarely need a flagship model, and the savings add up at volume.

Where can I check the current price? On Anthropic's own pricing page. Third-party reports vary, so treat them as a guide only.

Sources: launch coverage from DataNorth, Inc42 and BetaNews, September 2026. Pricing is as reported by third parties and has not been checked against Anthropic's own page.