Claude Mythos Preview got Anthropic's own AI running on AMD MI355X GPUs in about three days with almost no human help, according to Anthropic. The company had estimated six to twelve months for such a move, Tom's Hardware reported on November 16, 2026, citing the IPO roadshow.
Key facts
- According to Anthropic, Claude Mythos Preview got prompt processing for an Anthropic model running on AMD MI355X in about three days; the estimate was 6–12 months.
- Co-founder Tom Brown says the model wrote or ported almost 100% of the kernels with basically no human intervention.
- The claim comes from the IPO roadshow (video dated Nov 10, 2026); there is no independent test and no comparison with Nvidia.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
What Anthropic says about Claude Mythos and AMD
The figure comes from Anthropic co-founder and chief compute officer Tom Brown, in the roadshow video for the Anthropic IPO, dated November 10, 2026 according to Tom's Hardware. The key points:
- Time: about three days, against an internal estimate of six to twelve months.
- Work: Claude Mythos Preview ported or wrote almost 100% of the kernels, the low-level code that makes a model run fast on a chip.
- Scope: prompt processing for one Anthropic model on AMD MI355X, "at a useful level of performance".
The prospectus is more careful and only says new chip platforms can be onboarded in a fraction of the time it once took.
What the number does not say
This is material for selling shares, not an independent test. Tom's Hardware notes it is unclear whether generating responses was ported too, and how AMD's speed compares with Nvidia's. Anthropic already runs models on Nvidia, AMD, Google TPUs and AWS Trainium, and both Nvidia and AMD are buying shares in the IPO. So I read the three days as a direction, not a benchmark.
Why it matters beyond chips
Nvidia's strongest moat was never just the chip but the software around it: moving meant months of rewriting. That barrier is what shrinks here. As I see it, this is the same shift as the AI model becoming a swappable part, one layer down. Code is no longer a moat.
That reaches your marketing stack too. "Migration takes six months" is the quietest sales argument of every software vendor and agency. Once AI agents handle migrations in days, switching becomes a negotiation point. But only if you get your data out: an agent ports code, not exports your contract does not allow.
What I would do now
At every renewal (CRM, newsletter, analytics, agency), put two things in writing: a full data export in an open format and a short notice period. Then tell the vendor you have priced the move. How tightly Anthropic plans its own capacity is covered in Anthropic buys compute on the low side.
Got AI news? Submit news →
FAQ
What did Claude Mythos do with AMD?
According to Anthropic, Claude Mythos Preview ported almost all kernels so an Anthropic model processes prompts on AMD MI355X, in about three days instead of an estimated six to twelve months.
Is the three-day claim independently verified?
No. It comes from Anthropic's IPO roadshow. Tom's Hardware notes it is unclear whether response generation was ported and how AMD's speed compares with Nvidia.
Source
More sources
- Anthropic – Quelle der Zahlen laut Tom's Hardware; Video nicht selbst ausgewertet




