On September 22, 2026, Anthropic released Claude Opus 5.5, a new model that the company says matches its top-tier Claude Fable 5.1 on most work while costing less to run than the previous Opus model. Six days later, on September 28, 2026, it released the smaller Claude Sonnet 5.5. For businesses that run or plan AI features and AI agents, the practical news is lower running costs for long, multi-step tasks, plus a few API changes that existing integrations need to handle before switching.
What was announced
In its Opus 5.5 announcement, Anthropic describes the model as the first in a new Claude 5.5 family. The company states that it performs at the level of Claude Fable 5.1 on most work and costs 40 percent less to run than Claude Opus 5 on typical workloads. That figure is Anthropic's own estimate, and it combines lower prices with the model needing fewer steps and less output to finish a task.
The status of each part of the announcement is different, so it helps to separate them:
- Released and available now: Claude Opus 5.5 (September 22, 2026) and Claude Sonnet 5.5 (September 28, 2026). Anthropic's platform release notes list both on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.
- Research preview: a "fast mode" for Opus 5.5, which the announcement says is up to 2.5 times faster at a higher price.
- Beta: the ability to define new tools partway through a conversation, listed in the release notes as a beta feature.
- Announced, no date given: Claude Haiku 5.5, the low-cost model in the family. The Sonnet 5.5 announcement says only that it will arrive in the coming weeks.
Key details
The prices below are per million tokens (the units of text a model reads and writes) and come from the Opus 5.5 announcement and Anthropic's models overview. Check Anthropic's pricing page before budgeting, because cloud platforms can price differently.
| Item | Claude Opus 5.5 | Claude Opus 5 (previous) | Claude Sonnet 5.5 |
|---|---|---|---|
| Release date | September 22, 2026 | Still available as a legacy model | September 28, 2026 |
| Input price | USD 4 | USD 5 | USD 2 |
| Output price | USD 20 | USD 25 | USD 10 |
| Cached input (cache read) price | USD 0.20 | USD 0.50 | USD 0.20 |
| Context window | 1 million tokens | Not compared here | 1 million tokens |
| Maximum output | 128,000 tokens | Not compared here | 128,000 tokens |
Three further points from the same sources:
- Speed. Anthropic says Opus 5.5 generates output more than 30 percent faster than Opus 5, and makes a similar claim for Sonnet 5.5 against Sonnet 5.
- Where the saving comes from. List prices fell 20 percent, but cached input fell 60 percent. Anthropic notes that cache reads make up most of the cost of agent and coding workloads, because an agent re-reads the same instructions and history on every step.
- Restricted topics. The announcement says some cybersecurity requests are handled by an older model (Claude Opus 4.8) and that advanced biology work requires enrollment in a verification program. Most business applications will not notice, but security and life sciences products should read the details.
The benchmark results in the announcement are vendor-reported. They are a reason to test the model, not a substitute for testing it on your own tasks.
What this means for your business
This section is our interpretation, not part of Anthropic's announcement.
Agent projects that were too expensive may now be worth re-costing. An AI agent, meaning software that plans and carries out a multi-step task using tools, makes many model calls per job. The cost of one completed job matters more than the price of one call. A cut in cached input pricing lowers the cost of exactly that pattern. If you are unsure whether you need an agent at all, our guide to AI agents vs chatbots vs workflow automation explains the difference.
The choice between models is now a real design decision. Anthropic's documentation recommends starting with Opus 5.5 for most workloads and reserving Fable 5.1, which the models overview lists at USD 10 input and USD 50 output per million tokens, for the hardest reasoning. Sonnet 5.5 at half the Opus price suits well-defined, high-volume tasks such as classifying support tickets or drafting replies. Many products should route different tasks to different models.
Switching is not just a setting change. The release notes list breaking changes for Opus 5.5. If your software was built on an older Claude model, some requests that work today will return errors on the new one until a developer updates them.
US and UK readers: the model is offered through Anthropic directly and through the three large US cloud providers. Where your data is processed and stored depends on which platform and region you choose, so UK companies with data residency requirements should confirm this with the platform they use. Anthropic states that a zero data retention option is available for Opus 5.5. This is general information, not legal advice.
What to do now
- Find out which model versions your products call today. Anthropic's deprecations page explains how to export usage by model.
- Check retirement dates. The same page shows Claude Sonnet 4.5 was deprecated on September 30, 2026 and will be retired on November 30, 2026, after which requests to it will fail.
- Run your own test set on Opus 5.5 and Sonnet 5.5 and compare cost per completed task, accuracy and response time against your current model.
- Budget developer time for the breaking changes before switching production traffic.
- Do not plan around Haiku 5.5 or fast mode yet. One has no release date and the other is a research preview.
- Keep human approval and monitoring in place for agents that take actions. A more capable model does not remove the need for permissions and review.
For technical readers
Per the release notes, the Opus 5.5 model ID is claude-opus-5-5. Adaptive thinking is always on: requests that try to disable or manually enable thinking return a 400 error, and depth is controlled with the effort parameter instead. Forced tool use (tool_choice of type "any" or "tool") also returns a 400 error; the notes point to "auto" with strict tool use. Computer use requires the newer computer_toolset_20260801 on the Claude API and Google Cloud. Sonnet 5.5 has similar changes. Anthropic links a migration guide.
Frequently asked questions
Is Claude Opus 5.5 available now?
Yes. Anthropic released it on September 22, 2026, and its release notes list it on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Fast mode is a research preview, and Claude Haiku 5.5 has been announced without a date.
Will our AI costs automatically drop by 40 percent?
No. The 40 percent figure is Anthropic's estimate for typical workloads compared with Opus 5. List prices are 20 percent lower and cached input is 60 percent lower, so your saving depends on how much your application uses caching, how long its tasks are, and which model you are moving from. Measure it on your own workload.
Do we have to migrate from our current model?
Only if your current model is being retired. Opus 5 remains available, and Anthropic's deprecations page says it will not be retired before July 24, 2027. Claude Sonnet 4.5 users do need to move before November 30, 2026.
Conclusion
Claude Opus 5.5 is a meaningful release mainly because of cost: Anthropic has lowered the price of the long, repetitive model usage that AI agents depend on, and added a faster mid-tier model a week later. The sensible response is to test it against your own tasks, check retirement dates for whatever you run today, and plan the code changes before switching. If you would like help estimating what an AI feature or agent would involve for your product, you can request a quote or see our development services.
