Anthropic says Mythos got Claude running on AMD chips in three days, almost unaided

The claim sits in the IPO video Anthropic made for buyers, beside a promise that Claude will find its own efficiency gains.

By Think Facility · November 12, 2026 · Anthropic AI research

Part of a running story: Anthropic ipo (2 parts)

Anthropic's IPO roadshow started on November 10, 2026, and the company put a roadshow video on its IPO website for anyone thinking of buying shares. It runs in 11 chapters, from an introduction to a conclusion, with customer segments in between.

The chapter called Approach to compute carries a claim about Anthropic's own engineering. In it, Anthropic says its Mythos model got Claude running on AMD's chips in three days, a job it says used to take dozens of engineers months.

Roadshow starts: November 10th (EST)
The claim: Claude running on AMD in just three days
Who did the work: Mythos, with basically no human intervention
Kernels: Claude ported or wrote almost 100%
Price expected: November 18th (EST)
Trading expected: November 19th (EST)

What the video says about chips

Anthropic hasn't posted a transcript. Simon Willison made one with transcription software on November 11 and posted highlights. He names the speaker in that chapter as co-founder Tom Brown. Everything below from the video comes from Willison's transcript, so read the wording as close, not certified.

Brown says Anthropic now trains and serves Claude on four families of chips: AWS Trainium, Google's TPUs, Nvidia's GPUs and AMD's GPUs. Each one needs its own software stack, and he says many people once thought that wasn't worth the effort. AMD was new this year, and Anthropic handed the port to Mythos.

Andrew Curran posted the passage on X that evening, and his version matches Willison's almost word for word.

Andrew Curran@AndrewCurran_Nov 10, 8:09 pm ET

From the Anthropic Road Show video: 'This has also made us really good at bringing up new chips. As an example, this year, we started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work. This time we had Mythos try it out. It got Claude running on AMD in just 3 days with basically no human intervention. Claude ported or wrote almost 100% of the kernels.'

View the post on X

A kernel is the code that does the actual math on the chip. Writing them by hand for a new GPU is slow, specialist work, and it's the part Brown says Claude did nearly all of.

Claude finding its own savings

The chapter's second half is about what Anthropic calls compute multipliers, research gains that get more intelligence out of the same hardware. Brown's example: a 4x multiplier means a model that needed 100 units of compute can be trained on 25.

Then he goes further. He says this can speed up as Anthropic approaches recursive self-improvement, and that Claude "will be able to discover compute multipliers entirely by itself", so the gains keep compounding.

The factsheet Anthropic filed with the SEC on November 10 puts a number on how much of the company's code Claude already writes: more than 90% of contributions to its codebase in September 2026. Its summary line reads like this.

Each version of Claude helps build the next, continuously accelerating our pace of development and strengthening our leadership position

From anthropicpbc-fwpipowebsi

None of this has been tested by anyone outside Anthropic. The video doesn't say which AMD chips were used, how many people checked Mythos's work, or what "running" covered (training, serving, or both).

The line about cancer

The press picked up a different moment. In his closing pitch, Dario Amodei says cancer, Alzheimer's, heart disease and mental illness will be relics of the past in less than a decade, by Bearly AI's transcription posted on November 10. The founders' letter Anthropic filed the same day lists those diseases too, but softens it with "potentially" and gives no timeline in that sentence.

Why it matters to people buying in

This video is a pitch to ordinary investors as well as funds. The IPO website says anyone can ask for shares through a participating brokerage, that there's no guarantee of getting any, and that the price is expected on November 18th, with trading on Nasdaq under ANT the next day.

So a retail buyer weighing the stock is being told that Anthropic's AI already does a large share of its own engineering and will one day find efficiency gains alone. Those are the claims that would most change what the company is worth, and they're the ones nobody outside has checked. I'd want the prospectus to say the same thing in writing before betting on it.

It also says something about engineering work. Brown's own comparison is dozens of engineers for months against one model for three days. Earlier this month METR estimated that Anthropic's model progress had roughly doubled in pace, mostly from using its own AI. The roadshow video names one job and one number of days.

Shares are expected to begin trading on November 19th.