Anthropic has restricted access to Mythos on cybersecurity grounds, given what the company calls its ability to quickly identify vulnerabilities in critical infrastructure, such as banking, power grids, and software.
When Project Glasswing launched, some critics accused Anthropic of overhyping the threat so as to generate attention.
But companies that have tested Mythos have since endorsed its capabilities, with the US Government – which had been in a legal dispute with Anthropic – eventually testing the model over security concerns.
The White House has since set up an arrangement to test the most powerful models from the leading AI companies before they are released.
Anthropic says the restrictions on Fable 5 will work like this: most queries to it relating to sensitive issues like cybersecurity will be routed instead to the lower-tier model, Opus 4.8, which was made public in late May.
Anthropic hired outside experts to spend more than 1000 hours trying to find ways to bypass these restrictions – a process known as red-teaming.
The company also ran a bug bounty programme, which pays people to find security flaws. No one found a way to completely unlock the model, Anthropic said.
The security question, which Anthropic has made one of its main commercial arguments, had drawn the San Francisco startup into an unprecedented standoff with the Trump administration over its refusal to lift its restrictions on mass surveillance and autonomous lethal weapons.
After that dispute, the Pentagon severed its contracts with the company, whose AI tools had been the only ones to hold defence security clearance.
The launch of Fable 5 comes with a steep price tag – US$10 ($17) per million input tokens and US$50 per million output tokens, double the cost of Opus 4.8. Tokens are measures of usage of an AI model.
An intensive coding session by a programmer can burn through one million tokens in a matter of hours or less.
Despite exponential revenue growth, Anthropic remains far from profitability and is paying a premium for computing power. It recently began leasing a data centre from Elon Musk for US$1.25 billion per month.
These launches come amid intense financial excitement around AI.
Anthropic and rival OpenAI recently announced they had filed IPO plans, just days before a record-breaking market debut for SpaceX, Elon Musk’s space giant that also owns xAI, his AI company.
– AFP