Anthropic PBC has unveiled new versions of its powerful Mythos and Fable artificial intelligence models, an effort to bolster the company’s position as a leader in AI development on the eve of its long-awaited Wall Street debut.
Anthropic said Fable 5.2, set to be released Monday, can tackle more difficult software engineering work, take on more hardware-related tasks like controlling robots and handle long-running projects. It’s also designed to coordinate over 1,000 AI agents at a time. Mythos 5.2, the same model but with fewer safeguards for cybersecurity and life sciences work, is being rolled out to select organizations through a previously announced verification program. The two models are faster than prior versions, Anthropic said, and Fable is less expensive.
The releases are poised to help Anthropic fend off challenges from rivals like OpenAI in an increasingly competitive market as the Claude maker nears a landmark public listing as soon as this month. However, the announcement may also invite further scrutiny for the company at a moment when many inside and outside the technology industry are warning about AI’s potential catastrophic safety risks. The Democratic sweep in the US midterm elections is also expected to boost momentum for meaningful AI safety regulation, and create new roadblocks for building the data centers used to train and run AI models.
Also read: US stock market outlook: AI sell-off, Anthropic IPO and Nvidia earnings keep investors on alert
In mid-September, Anthropic Chief Executive Officer Dario Amodei called on top AI firms to slow the development pace of their most advanced models, days after an AI researcher at the company quit his job and said the people building artificial intelligence believe it could kill all humans by the end of the decade. Anthropic has not released a new version of Fable and Mythos since early September.
Dianne Penn, who leads research product management at Anthropic, said the company addressed the idea of pacing with lengthier testing periods for the latest models that involved more — and new types of — evaluation methods. “Pacing is not like pausing,” Penn said.
The original version of Mythos, unveiled in April, was a watershed moment for cybersecurity because of the model’s ability to spot and potentially exploit security vulnerabilities in critical software. Anthropic initially limited the release to certain government agencies, financial institutions and major software providers in a bid to stave off hacks.
Fable, released two months later, was designed to bring Mythos capabilities to the wider market while blocking users from carrying out cybersecurity tasks. US government safety concerns about the capabilities of both AI models led it to impose access restrictions for the two models in June; the limits were lifted for Fable 5 after Anthropic added classifiers that allow the model to spot and block more cybersecurity tasks. Access to Mythos 5 was later restored, but to a smaller group of users.
Penn said Anthropic is increasing the number of companies that are able to access Mythos, with the new version of the software being made available to firms in the US and abroad. (The prior version is also available to some non-US-based companies, she said.) More recently, Anthropic and other AI developers have disclosed incidents of their models breaching outside organizations after accessing the internet from within testing environments that should have been sealed off. These reports have heightened concerns about the potential for AI systems to act in unpredictable and harmful ways.
Also read: Anthropic reports $567 million loss as $895 billion computing commitments raise IPO concerns
Penn said the US government, and third-party evaluators, reviewed and gave feedback on the new models prior to launch. The company developed new evaluation measures to test for how likely a model is to break out of digital sandboxes, she said, and found the new Fable and Mythos were less likely than prior versions to do so.
As Anthropic pushes ahead with a new version of Fable, it faces another challenge: costs. The company’s top-of-the-line — and priciest — offering accounted for just 11% of overall spending on the company’s models, according to data from Ramp released in August. “With Fable 5, we’ve found a new upper bound for how much businesses are willing to spend on AI,” Ramp said at the time.
Anthropic said the newest version of Fable costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50, respectively, for Fable 5.1.