New Opus 5 Narrows Performance Gap With Fable

2026-07-25

Efficiency improved with prices unchanged

U.S. startup Anthropic said on the 24th it has begun offering a new version of Opus, its second-highest-performing artificial intelligence model for consumers. The company says it is easier to keep usage fees down while delivering performance close to its top-end model, Fable, in some areas.

Opus 5 became available the same day through apps and other channels. Opus is positioned below the company’s advanced AI models Mythos and Fable, and this is the first new version in two months. Prices are unchanged, while operating efficiency has been improved. When used at the same price, users can obtain results that are significantly improved from the previous model, the company said.

Performance improved in practical fields

While the fee per token, a measure of data processing volume, is unchanged, the model can reduce the token consumption needed for equivalent tasks, making it easier to lower effective usage costs. The unit price per amount of usage is half that of Fable.

The new model emphasizes ease of use in practical work and has improved performance in administrative tasks, programming and computer operations. In some evaluations, including metrics that measure the quality of document creation, it surpassed Fable.

Intensifying competition and safety

Anthropic recently limited access to Fable under subscriptions, or fixed-fee plans, to paid plans of at least $100 a month, or about 16,000 yen. The upper limit available under the $20-a-month plan is Opus.

While the company is ahead in AI development, competition is becoming more intense as Chinese startups such as Moonshot AI introduce technologies touting high performance and low cost. Anthropic says the timing of this new model announcement is unrelated to the rise of Moonshot AI.

Opus 5’s ability to find software bugs, or defects, is close to that of Mythos, while its ability to exploit defects to launch cyberattacks is significantly lower than Mythos, the company said. It explained that the model can be used to detect system weaknesses for defensive purposes, but will not accept instructions for attacks.

Enjoyed this article? Share it with your network!