Anthropic put a new artificial intelligence (AI) model online on Friday, Claude Opus 5, presented as less susceptible to abuse and more difficult to hijack for computer attacks.
• Also read: Data center electricity consumption could quadruple by 2035
• Also read: How much of the music on Deezer is of AI origin? More than half!
In the Claude family, Opus 5 succeeds Opus 4.8, released at the end of May. Since then, Anthropic has also marketed Fable 5, a model of a new genre.
Fable 5 is a constrained version of Mythos 5, considered Anthropic’s most powerful AI and only released to a few companies and organizations for the purposes of securing the internet.
According to Anthropic, Opus 5 “approaches the capabilities of Fable 5 in many areas, for half the price,” according to a press release.
While Fable 5’s performance has been praised since its release in June, many have also noted its high cost compared to the competition, including the closest OpenAI model, GPT-5.6.
Long relegated to the background, due to generous subscriptions often offered at a loss, the question of the price of using AI has become central since the start of the year.
This is mainly due to the emergence of AI agents, interfaces capable of carrying out several complete tasks on request in everyday language, but also of testing the results and making modifications.
These agents consume significant computing capacity, which increases their cost compared to traditional generative AI.
Anthropic also highlighted the calibration of the model, announced as the “most aligned” of all the versions launched by the San Francisco company.
Alignment is the constraints imposed on an AI so that it behaves as its designers intended and does not deviate from certain principles.
The subject was put back at the center of discussions after the incident experienced by OpenAI, two models currently being tested left the confined environment in which they were tested to attack the Hugging Face AI platform.
Many specialists have been concerned about seeing AI take initiatives that contravene their instructions, moreover without informing their developers.
According to tests carried out by Anthropic, Opus 5 “displays the lowest rate of hidden behavior” among the most advanced models in the group.
Anthropic also emphasizes having “intentionally avoided training Opus 5 for cybersecurity tasks”, even if the general improvement of the model compared to its predecessors “has made it progress in this area”.
It “is close to the level of Mythos 5 in detecting computer vulnerabilities, but it is substantially behind Mythos 5 in terms of exploiting them” with a view to a possible computer attack, assures the group.
