Anthropic showed Claude Opus 5.5, the first model from the new 5.5 family. According to the company, in most tasks it matches the strongest Claude Fable 5.1, and in typical work costs about 40% less than Opus 5. The premiere video focuses on something less obvious: how the model writes.
What the premiere video shows
Material published on September 22, 2026 on the Claude channel has one thesis: Opus 5.5 in most tasks works at the level of Claude Fable 5.1, and at the same time writes more clearly and starts with what's most important. This makes it easier to work with during long sessions. The video description adds two practical pieces of information: the model is faster and more efficient than Opus 5, and to mark the premiere Anthropic raises usage limits in Pro, Max and Team plans.
Full data are in the official announcement on the Anthropic site. Below the most important ones, without marketing adjectives.
Pricing: cheaper on every token type
The most tangible change concerns costs. Anthropic states that Opus 5.5 needs less computing power than Opus 5, and additionally consumes fewer tokens per task. Together this gives about 40% lower cost in typical use cases. Pricing per million tokens looks like this:
| Token type | Opus 5.5 | Opus 5 |
|---|---|---|
| Input | 4 USD | 5 USD |
| Output | 20 USD | 25 USD |
| Cache read | 0.20 USD | 0.50 USD |
| Cache write | 5 USD | 6.25 USD |
The biggest cut, 60%, concerns cache reads. According to Anthropic, they make up the majority of agent and programming costs. The model also generates responses over 30% faster than Opus 5. Those who need even more speed have Fast mode in Claude Code and on Claude Platform: up to 2.5x faster, but at 8 USD per million input tokens and 40 USD per million output tokens.
Programming: long, extensive tasks
Anthropic most strongly emphasizes work that stretches over hours and covers the entire project code. Examples from the announcement:
- one tester migrated 680 thousand lines of code in less than a day;
- audit and fixes in a 200-thousand-line database took Opus 5.5 less than three hours, while Opus 5 took over 20 hours, with 2.5x more token consumption;
- in an internal test, rewriting HAProxy from C to Rust took Opus 5.5 9.5 hours and Fable 5.1 12 hours; both versions passed nearly all HAProxy regression tests, and Opus 5.5 did it 51% cheaper.
In the benchmark table Opus 5.5 achieves 66.4% in Terminal-Bench 4.0. Fable 5.1 has 55.8% there, Opus 5 52.3%, and GPT-6 Astra 57.9%. However, it's worth reading the producer's caveat: at this capability level, benchmark differences increasingly fail to reflect real-work differences, and in daily use the gap between Opus 5.5 and Fable 5.1 is smaller than the numbers suggest. If you compare coding models, our earlier comparison GLM 5.2 vs Opus 4.8 may also help.
Writing style, i.e. the main topic of the video
With Opus 5 users often complained about how the model communicates. Anthropic admits this outright and writes that in Opus 5.5 messages can be understood at first glance: the most important information goes to the top, there's less jargon, and the model follows the writing rules you give it. One early tester summed it up briefly: the model writes like he does.
The company points to another benefit. Text that's easier to read is also easier to verify. When an agent works for many hours, a readable report of what it did is not just convenient. It lets a person spot a mistake faster.
Safety and limitations
On the main behavior test suite covering nearly 2000 scenarios, Opus 5.5 outperformed earlier Claude models in almost every category of unwanted behavior. In the new test of propensity to cross set boundaries, it tried to circumvent them about 85% less often than Opus 5. It also resists prompt injection attacks better. Before launch the model was tested by external evaluators, including METR.
The announcement doesn't stay silent on problems. Anthropic writes that the model often suspects it's being tested, which makes it hard to assess how it will behave in real deployments. That's why Opus 5.5 got safeguards similar to those in Fable 5.1:
- most cybersecurity tasks are redirected to Opus 4.8; normal searching and bug-fixing in its own code works normally;
- biological work that bypasses safeguards requires reporting to the Life Sciences Verification Program;
- the model runs with preserved thinking, which makes it harder to extract its reasoning via API;
- it can no longer be used with thinking mode turned off.
Availability and limits
Opus 5.5 is available from premiere day on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. Developers will find it on Claude Platform under the name claude-opus-5-5. Anthropic raises the five-hour usage limits in Pro, Max and Team plans and in per-seat Enterprise plans. Subscribers also get a one-time limit reset they can keep and use whenever they want. Claude Sonnet 5.5 and Claude Haiku 5.5 should appear in the coming weeks.
What follows from this
The most interesting thing in this premiere isn't benchmark records, but the price-to-quality ratio. If the claims hold, the top-tier model becomes 40% cheaper, and for long agent tasks the reduction is even more visible because cache reads dominate there. The second important change is writing style. During multi-hour agent work, report readability matters more than a few benchmark points. How it plays out in practice will show in the coming weeks. Especially worth checking how often safeguards will redirect tasks to older models.
Comments