Anthropic released Claude Opus 5.5 on September 22, 2026. The promise is quickly told: On most tasks, the new model works at the level of Fable 5.1, but costs around 40 percent less than Opus 5 in typical use. As a little bonus on top: All subscriptions get a one-time usage limit reset for playing around with Opus 5.5. It can be redeemed until October 22 and, with clever timing, delivers more than you might think at first glance.
Classification: the second Opus model of the fifth generation
Anthropic describes Opus 5.5 as the first model of the new Claude 5.5 family. Claude Sonnet 5.5 and Claude Haiku 5.5 are expected to follow in the coming weeks. Within the Opus line, it is the second model of the fifth generation, following Opus 5 from July 2026. Accordingly, the system card describes it as an upgrade to Opus 5.
The benchmark there is set by the Mythos models, whose safeguarded variant is generally available under the name Fable. I described the stir Fable 5 and Mythos 5 caused in June, when a US government order forced them to be shut down, in an earlier article on model dependency. In many evaluations, Opus 5.5 matches or exceeds both Fable 5.1 and Mythos 5.1.
Usage reset until October 22: choosing the right moment
With the launch, Anthropic is raising the five-hour limits on the Pro, Max and Team plans as well as on the seat-based Enterprise plan. In addition, every subscription gets a one-time usage limit reset that you can save and redeem until October 22, 2026. You redeem it under Settings > Usage, in the browser or in Claude Desktop. This is not possible in the mobile app or in Claude Code in the terminal. A redeemed reset cannot be undone, and it expires if you downgrade or cancel.
The reset restores your entire allowance, including the separately limited allowance for Fable 5.1. That makes it worth waiting before you redeem it. Because Opus 5.5 works so economically, the regular allowance lasts considerably longer anyway. If you use the reset as soon as the general limit runs low, you may be giving away the Fable share. It is better to wait until the Fable allowance is largely used up as well and then reset both at once. Once again, Anthropic is being generous here.
Noticeable in usage: pricing and cost efficiency of Opus 5.5
The savings come from two directions. First, prices per million tokens are going down:
- Input: $4 (Opus 5: $5)
- Output: $20 (Opus 5: $25)
- Cache reads: $0.20 (Opus 5: $0.50)
- Cache writes: $5 (Opus 5: $6.25)
For comparison: Fable 5.1 costs $10 for input and $50 for output. So Opus 5.5 delivers results close to Fable for less than half the token price.
Second, the model needs fewer tokens for the same task. According to Anthropic, most of the cost in agentic work and coding comes from cache reads, and this item drops by 60 percent. Together, this adds up to savings of around 40 percent on typical workloads. Output is also generated more than 30 percent faster than with Opus 5. If you need more speed, Claude Code and the Claude Platform offer a fast mode with up to 2.5 times the speed, though at twice the price.
What Opus 5.5 delivers: the key features
Long, branching programming tasks. Opus 5.5 shows its greatest strength in work that spans an entire codebase, such as migrations, refactorings and audits. It stays on task over long sessions and loses the thread less often. On Terminal-Bench 4.0, which measures multi-step tasks on the command line, it scores 66.4 percent, compared with 55.8 percent for Fable 5.1.
Research with reliable evidence. In an internal test, Anthropic had quarterly reports researched and automatically checked every figure and every quote against the sources. A single invented detail meant failing. Opus 5.5 passed in 16 out of 18 cases, while Fable 5.1 and Opus 5 passed in none. For me, this is the most interesting number in the entire announcement.
Clearer communication. Opus 5.5 puts the most important information first, is less likely to use jargon and follows the writing rules you give it. In long working sessions, this makes the results easier to follow and check.
More context and output. The context window holds one million tokens, and a single response can be up to 128,000 tokens. Thinking adapts to the task and is set to medium intensity by default.
Safety. In the automated behavioral audit with nearly 2,000 scenarios, Opus 5.5 performs better than any Claude model tested so far. In a new test, it attempted to get around the boundaries set for it around 85 percent less often than Opus 5, and it is at least as robust against prompt injection as its predecessor.
Where the numbers reach their limits and what counts in everyday work
Anthropic is cautious about its own benchmarks. At this level of performance, it says, gaps in tests are a less reliable indicator of differences in everyday use, and in its own use the gap to Fable 5.1 is smaller than the scores suggest. For particularly demanding and long reasoning tasks, Anthropic still recommends Fable 5.1, and for some task types the competition is ahead.
My own impression fits this assessment. I switched from Opus 5 to Opus 5.5 in the middle of a running session, and the transition was seamless and barely noticeable at first. That is partly because Opus 5 was already a strong model. Opus 5.5 works reliably and confidently, and the results come across as high-quality and plausible. What I clearly notice is the lower consumption, even on complex, long tasks.
For choosing a model, the ranking therefore remains a starting point. A decision only becomes robust with your own reference tasks that measure quality, runtime and cost per task together. The usage reset is a good opportunity to set up such a comparison properly.
Safeguards and what changes for integrations
Opus 5.5 is the first Opus model to launch with safeguards similar to those of Fable 5.1, covering cybersecurity, biology and distillation. Finding and fixing bugs in your own code works as usual in everyday development. However, Anthropic transparently routes most cybersecurity tasks to Claude Opus 4.8. Vetted organizations can apply to separate verification programs for security and biology work.
If you connect Opus 5.5 via the API, test before switching. Thinking mode can no longer be turned off, forced tool calls return an error, and thinking blocks are tied to the model and the conversation. For API accounts created on or after August 31, 2026, preserved thinking also applies, which prevents subsequent edits to earlier reasoning steps. Outputs carry watermarks in line with the EU AI Act, and as with earlier Opus models, zero data retention is available. Opus 5.5 is available on the Pro, Max, Team and Enterprise plans as well as via the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. In the API, it is called claude-opus-5-5.
My takeaway after the first few days with Opus 5.5
For me, Opus 5.5 came at the right time. Over the weekend, I worked on several development projects in parallel and used up my Max plan allowance earlier than I have in a long time. That makes me all the happier about the small mid-week boost. It lets me step it up once more, instead of having to slow down and work cautiously. And after the first few days, I can say: The announced savings are clearly noticeable in usage, and my allowance lasts noticeably longer for comparable work.
What will you use the usage reset for, and which task will you pick to put Opus 5.5 up against your current model?
Image: Anthropic
