Claude Opus 5.5 Release: 20% Cheaper, Thinking Can't Be Turned Off
Claude Opus 5.5 is a model Anthropic released on September 22, 2026. Its API model ID is claude-opus-5-5. Two changes matter most to developers: it costs 20% less than Opus 5, and thinking can no longer be turned off. If your code explicitly disables thinking, or relies on an older combination of parameters, check it before you switch models. This article draws on Anthropic's official pages as of October 2026; the official documentation is the final word.
What did Anthropic announce for Opus 5.5?
According to the official Anthropic release page:
| Item | Official statement |
|---|---|
| Release date | September 22, 2026 |
| Model ID | claude-opus-5-5 |
| Input / output | $4 / $20 per million tokens, 20% lower than Opus 5 |
| Cache reads | $0.20 per million tokens, 60% lower than Opus 5 |
| Fast mode | Available in Claude Code and the Claude Platform, up to 2.5x faster, $8 / $40 per million tokens |
| Platforms | Launched simultaneously on AWS, Google Cloud and Microsoft Azure, with zero data retention supported |
Anthropic says that with default settings, a typical workload costs about 40% less than on Opus 5, with output more than 30% faster. That is the vendor's own claim. Whether it holds for your tasks is something you need to measure yourself.
What does "thinking can't be turned off" mean?
Anthropic states that thinking can no longer be disabled on Opus 5.5. If a request still carries a "disable thinking" setting, it may be rejected or ignored. Check the API documentation for the exact behavior and parameter names; this article does not make that call on Anthropic's behalf.
Alongside this comes the effort setting, with five levels: Low, Med, High, Xhigh and Max. The default is Medium. Since thinking can't be switched off, effort is now the main lever for controlling cost and latency: use a low level for simple tasks and save the higher levels for complex refactors.
What is preserved thinking, and who will run into it?
According to Anthropic, API accounts created on or after August 31, 2026 have preserved thinking enabled, as an anti-distillation measure. If your application rewrites message history across multi-turn conversations, especially the parts that contain thinking content, check the official documentation first to confirm which modifications are allowed.
What changed in Claude Code?
According to the Claude Code changelog, starting with version 2.1.280, Opus 5.5 is the default Opus in Claude Code, with a 1M-token context window. The default model for Pro and Team Standard plans also changed from Sonnet to Opus. A previously saved effort setting does not carry over to the new model, so after upgrading, confirm it in /model. (The changelog entries are long; refer to the official notes for version details.)
If you use Claude Code through a custom endpoint, see Configuring custom API endpoints for Claude Code, Codex and OpenCode.
What are the four steps to migrate to Opus 5.5?
- Change the model ID. Replace the old ID in your config with
claude-opus-5-5. Run it in a test environment first; don't cut production over all at once. - Check thinking-related parameters. Search your code for every place that disables thinking or sets a thinking budget, and verify each one against the official API documentation.
- Choose effort per task. Run the same batch of real tasks at Low, Medium and High, record quality, latency and token usage, and then decide your default.
- Re-estimate cost. Multiply the prices in the table above by your actual usage. If you have long prompt prefixes, the drop in cache-read pricing matters even more; see How to use Claude prompt caching to save money for how to set it up.
A header like this works for a comparison log:
Task type | effort | input tokens | output tokens | latency | usable result?
How do you keep an eye on it after launch?
- Model monitoring: based on OpenRouter's public API, it shows availability, latency and quantization changes across inference providers.
- Official site changes: monitors additions and removals on official pricing pages and changelogs. For the method, see How to find out the moment an LLM API price changes.
- Anthropic API status: cdnping probes the official endpoints on a schedule from Cloudflare's edge. cdnping started collecting data on 2026-10-17, so the history is still short and only shows recent trends.
For how to read availability and latency, see Three new Anthropic models in September and October: how to read availability and latency after launch.
Should you switch right away?
If you are on Opus 5 today, a 20% lower unit price per the official numbers justifies scheduling an evaluation. If your pipeline relies on turning thinking off to keep latency down, first test whether the Low level is good enough. Without a strong need, there is no reason to move production in the first week after release; roll it out to a small slice of traffic first.
