Anthropic replaced its flagship AI model on September 22, 2026, just two months after the previous version shipped. Claude Opus 5.5 is faster, cheaper, and, according to the company, harder to manipulate than anything it has released before.
The headline number is price. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens through the API. That is a 20% cut from Opus 5’s $5 and $25 rates, but Anthropic says typical workloads end up costing about 40% less, because the new model also needs fewer tokens to finish the same tasks.
What Opus 5.5 costs
API pricing per million tokens, compared with the model it replaces:
| Claude Opus 5.5 | Claude Opus 5 | |
|---|---|---|
| Input | $4.00 | $5.00 |
| Output | $20.00 | $25.00 |
| Cache read | $0.20 | $0.50 |
| Cache write | $5.00 | $6.25 |
For comparison, Claude Fable 5.1, Anthropic’s higher-tier model, is priced at $10 per million input tokens and $50 per million output tokens, which makes it 150% more expensive than Opus 5.5 at list price.
A new fast mode is also available in Claude Code and on the Claude platform, priced at $8 per million input tokens and $40 per million output tokens, with speeds up to 2.5 times faster than the standard mode.
Benchmarks: built for agents
Anthropic is positioning Opus 5.5 for long-running agentic coding and professional knowledge work, and the benchmark sheet reflects that. The company reports 66.4% on Terminal-Bench 4.0, 89.9% on SWE-bench Pro, 81.8% on OSWorld 2.0 for computer use, and 1,846 Elo on GDPval-AA v2.1. On Humanity’s Last Exam, a broad reasoning test, it scores 67.7%.
Third-party tracking puts those numbers in context. GPT-6 Astra scores 57.9% on Terminal-Bench 4.0 and Claude Fable 5.1 scores 55.8%, against 52.3% for the old Opus 5. In other words, Opus 5.5 matches or beats the pricier Fable 5.1 while costing less than half as much per token.
Speed matters as much as scores. Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5, which compounds across agentic workflows that chain dozens of model calls together. In company-run trials, the model completed a software migration of around 680,000 lines of code in less than a day.
One caveat is worth noting. Testing by Artificial Analysis found that Opus 5.5 consumes considerably more tokens per task than some rivals, about 119,000 tokens on a standard test set versus 17,000 for OpenAI’s GPT-6 Astra. Cheaper per token does not always mean cheaper per task, so heavy users should measure their own workloads before switching.
A price war broke out the same day
Opus 5.5 did not launch into a quiet market. On the same day, September 22, OpenAI released GPT-6 Sol at $2 per million input tokens and $10 per million output, alongside the budget GPT-6 Luna at $0.10 input and $0.50 output. Two frontier labs shipping flagship models on the same afternoon, both slashing prices, is not a coincidence. It is a fight over the cost of useful work.
That framing is the real story of this release cycle. Vendors are no longer competing purely on benchmark scores. They are competing on how much work a dollar buys, which is why Anthropic leads with operating cost rather than a single benchmark number. Readers who want background on OpenAI’s chatbot strategy can start with our earlier explainer.
Checked by outsiders, not just Anthropic
Anthropic had Opus 5.5 evaluated by outside groups before launch, including METR and Frontier Design. Reporting on those evaluations notes that boundary-circumvention attempts fell 85% compared with Opus 5 and Mythos 5.1. The company describes it as the strongest alignment result it has shipped to date.
Availability and what comes next
Opus 5.5 is live on the Claude platform and API, plus Amazon Bedrock, Google Cloud, and Microsoft Azure, with a zero-data-retention option for regulated customers. It carries a 1 million token context window, up to 128K tokens of output per request, and a June 2026 knowledge cutoff.
Anthropic also raised the five-hour usage limits on Claude subscription plans, answering one of the most common complaints from power users. Claude Sonnet 5.5 and Claude Haiku 5.5 are expected to follow in the coming weeks.
The takeaway for buyers is simple. Frontier capability now arrives with falling prices attached, and the gap between flagship and second-tier models is narrowing fast. If your AI budget was set six months ago, it is probably wrong now. For a sense of where all this capability is heading, see our roundup of the hottest AI applications.
Sources: Anthropic; VentureBeat; Gadgets360; Artificial Analysis.