Anthropic released Claude Opus 5.5 on 22 September, and within hours it had taken first place on the Artificial Analysis Intelligence Index from Claude Fable 5.1, Anthropic’s own previous leader. The company now holds the first three positions on that index.
It also did something the frontier labs rarely do with a flagship model. It made it cheaper. Input tokens cost $5.60 per million and output tokens $28, down from $7 and $35 on Claude Opus 5, according to Anthropic. For Australian teams running AI inside real workflows, a 20% cut on the best model available changes the arithmetic.
What Anthropic actually changed
The headline numbers are a 20% reduction on both input and output pricing. Cache reads fall further, down 60% to $0.28 per million tokens, and cache writes drop to $7 from $8.80. Anthropic bills in United States dollars, so what an Australian team pays each month moves with the exchange rate.
Speed moved too. Anthropic says output generates more than 30% faster than Opus 5, and a separate fast mode runs up to two and a half times quicker at $11 per million input tokens and $56 per million output. The company puts the combined effect at 40% less to run than Opus 5 on typical workloads.
Availability is immediate rather than staged. Anthropic listed the model on its own platform alongside Amazon Web Services, Google Cloud and Microsoft Azure on day one, and opened it to Pro, Max, Team and Enterprise subscribers. Buyers already committed to one of the three major clouds do not have to move anything to test it.
Where it lands on the benchmarks
On Anthropic’s own testing, Claude Opus 5.5 scored 66.4% on Terminal Bench 4.0 against 55.8% for Claude Fable 5.1, and 54.4% on FrontierCode against 50.3%. It also posted 57.8% on CursorBench 4.0, 81.8% on OSWorld 2.0 and 1846 Elo on GDPval.
Vendor benchmarks deserve caution, so the useful signal came from elsewhere. The Artificial Analysis index scored Claude Opus 5.5 at 58 against 53 for Claude Fable 5.1, which had held the top position until this release.
The cost column is where it gets interesting for buyers. The third placed configuration scores 54 at roughly $2.56 per task, against $4.58 for GPT 6 Astra at its highest setting, which sits sixth with 53. Four models dropped out of the top 20 entirely, including MiMo V2.6 Pro and Qwen3.8 Max.
The churn underneath tells its own story. Four models fell out of the top 20 on the same update, including MiMo V2.6 Pro from Xiaomi and Qwen3.8 Max, both of which had been there a week earlier. The gap between the frontier and the fast following open weight models widened rather than closed.

The safety numbers matter for agents
Anthropic reports that Claude Opus 5.5 achieved the best scores of any of its models on an automated behavioural audit, and that it is much less likely to take actions that are hard to reverse. It also describes the model as more resistant to prompt injection than Opus 5.
That last claim carries weight for anyone letting a model touch live systems. An agent that files invoices, updates a CRM or processes contracts is exposed to whatever text arrives in those documents, and prompt injection is the practical route into that exposure. Teams building document automation have been sizing that risk for two years without a clean number to point at.
Anthropic has not published the audit methodology in full, so the claim sits somewhere between a benchmark and a statement of intent. It is still the first time the company has led a launch with behavioural numbers rather than capability ones, which says something about what enterprise buyers have been asking for.
Why Australian businesses should care
Token pricing rarely decides whether a project goes ahead, but it decides how far it scales. A support workflow or a document pipeline that was affordable across one department at $35 per million output tokens can now run across three at $28, and the cache read cut of 60% lands hardest on exactly the workloads that resend the same context all day.
A worked example helps. A support assistant handling 2,000 conversations a month at roughly 4,000 output tokens each would have cost about $280 a month in output charges on Opus 5. The same volume now sits near $224, before the cache savings land on the system prompt that goes out with every single one of those conversations.
The data residency question is separate and still unresolved for many buyers. Anthropic committed to a Queensland data centre earlier this month, which we covered in our report on the Queensland announcement, and organisations working under the Privacy Act 1988 will want that live before moving sensitive workloads.
For teams already comparing the two Anthropic flagships, the ground has shifted rather than settled. Our breakdown of Opus 5 against Fable 5 now has a third model sitting above both on price and on measured capability.
What this means for the model race
Price cuts on frontier models used to arrive when something better shipped and the old model was pushed down the shelf. This one arrived attached to the new leader, which suggests the cost of serving these models is falling faster than the labs have been passing on.
OpenAI added two model identifiers to its documentation the same evening, so the top of the table is unlikely to hold for long. What has changed is the floor. A model at the top of an independent index at under $3 per task is a different proposition to Australian mid market buyers than anything available a month ago.
The teams that benefit are the ones with workloads already built and instrumented, because they can swap the model and read the difference in a week. Everyone else is still deciding what to automate, and for them the price of the model was never the binding constraint.
