Nabugu - stock.adobe.com

News

Anthropic unveils Opus 5.5: powerful performance, still premium price

The AI lab still lags in the price war with rivals.

Esther Shittu
By
Published: 22 Sep 2026

Anthropic on Tuesday introduced Claude Opus 5.5, the first model in the Claude 5.5 family.

The vendor is trying to compete on price while appearing to follow its own principles, after CEO Dario Amodei called for a three-step plan to slow AI development by building the technology at a more restrained pace.

Amodei authored a widely read Sept. 13 blog post asserting that to moderate the development of AI technology, AI companies should allow third-party evaluators to examine the safety provisions of their models. Amodei also called for coordinating industry safety standards among vendors and the U.S. and other democratic governments.

Claude Opus 5.5 is good at advanced agentic coding, specializing in long, complex, multi-step software tasks and optimized for heavy knowledge workloads such as running financial research, complex document analysis and comprehensive writing tasks, Anthropic said. The model includes a feature that enables it to scale its reasoning based on the prompt it receives.

Opus 5.5 differs from Claude 5 in pricing: the model costs $4 per million input tokens and $20 per million output tokens.

Comparatively, Opus 5 goes for $5 per million input tokens and $25 per million output tokens.

But despite Anthropic slashing the model's price by 20%, Opus 5.5 is still more expensive than rival lab OpenAI’s GPT-6 Sol and GPT-6 Luna models, also released on Tuesday. GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 per million input tokens and $1.50 per million output tokens. The price of Opus 5.5 is a bit lower than OpenAI’s GPT-Astra, released on Sept. 3, which costs $10 per million input tokens and $50 per million output tokens.

The real pricing

Nevertheless, reducing token prices may well not make much of an impact for users, said Hendrik Krack, developer advocate at CodeRabbit, an AI software development and change management vendor that automates code reviews for engineering teams.

“A decrease in price, for most consumers, might not make that huge a difference, or it could even potentially become more expensive as the model tends to think much more and then output many more tokens as well, which you are ultimately going to be charged for by a model provider,” Krack said.

CodeRabbit conducted several evaluations of the model and found that the number of tokens Claude Opus 5.5 produced increased by 50%, even though the model received the same input as in previous model generations, he noted.

Despite the apparently ineffective price cut, CodeRabbit also found that the model performed impressively on long-horizon coding tasks after being evaluated for 24 to 48 hours.

“The sort of things that it builds … really surprised our team,” Krack said.

In complex bug detection tasks, Krack said Claude Opus 5.5 found twice as many bugs as CodeRabbit’s current baseline model. With standard problems, though, there appeared to be little difference between the baseline model and Opus 5.5, he said.

Krack noted that the model also requires highly specific instructions to achieve a goal.

”You have to be sure that these instructions that you put in are very explicit and very well formulated,” he said.

A shift

With Opus 5.5, Anthropic also seems to understand that the frontier model battle is changing, said Sid Nag, founder of Tekonyx, an independent AI analyst firm.

“They’re recognizing the fact that the model race is shifting from not just technology, maximum intelligence, maximum reasoning,” Nag said. “They are shifting to usable intelligence per dollar. It is a very financial shift.”

He added that Anthropic also appears to be adhering to its mission of slowing AI development while continuing to innovate.

The vendor claimed that Opus 5.5 is less likely to escape boundaries than Opus 5 or Claude Mythos 5.1. -- a significant characteristic, if confirmed, in light of recent alarming incidents in which Anthropic and OpenAI AI agents escaped their sandbox environments. Anthropic also said it is applying cybersecurity safeguards to the model, similar to those in its Fable 5.1 cyber model, and that it will soon include the latest model in its Cyber Verification Program.

“Anthropic is not just investing in risk prevention but pacing the rate of new capabilities so that risk prevention has time to catch up,” Nag said.

Esther Shittu is an Informa TechTarget news writer and podcast host covering AI software and systems.

Dig Deeper on AI Technologies & Platforms