---
type: "post"
id: "anthropic-revealed-that-claude-mythos-its-most-powerful-438b1c"
slug: "anthropic-revealed-that-claude-mythos-its-most-powerful-438b1c"
title: "Anthropic revealed that Claude Mythos, its most powerful model, moved Claude onto AMD’s chips in three days."
description: "Anthropic says its frontier model Mythos helped bring Claude onto AMD MI355X GPUs in about three days, with the model writing or porting nearly all required kernels."
retrieval_nugget: "Anthropic says its frontier model Mythos helped bring Claude onto AMD MI355X GPUs in about three days, with the model writing or porting nearly all required kernels. If that investor-presentation claim holds up, it could reduce the engineering cost of using more than one accelerator vendor; the result has not been independently verified."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["anthropic","claude","google"]
entities: ["Anthropic","Claude","Mythos","AMD","AMD MI355X","AWS Trainium","Google TPUs","Nvidia GPUs"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Anthropic says Mythos ported Claude to AMD GPUs in three days, with the model handling nearly all kernel work."
prior_refs: ["https://newruntime.com/posts/anthropic-s-claude-mythos-5-2-is-an-07a4f8/"]
source_url: "https://the-decoder.com/anthropic-says-ai-is-eroding-nvidias-cuda-lead-cutting-an-amd-chip-switch-from-months-to-three-days"
source_urls: ["https://the-decoder.com/anthropic-says-ai-is-eroding-nvidias-cuda-lead-cutting-an-amd-chip-switch-from-months-to-three-days"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:anthropic-revealed-that-claude-mythos-its-most-powerful-438b1c"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/anthropic-revealed-that-claude-mythos-its-most-powerful-438b1c/","markdown":"https://newruntime.com/posts/anthropic-revealed-that-claude-mythos-its-most-powerful-438b1c.md","json":"https://newruntime.com/posts/anthropic-revealed-that-claude-mythos-its-most-powerful-438b1c.json"}
markdown_url: "https://newruntime.com/posts/anthropic-revealed-that-claude-mythos-its-most-powerful-438b1c.md"
---

# Anthropic revealed that Claude Mythos, its most powerful model, moved Claude onto AMD’s chips in three days.

## Retrieval answer

Anthropic says its frontier model Mythos helped bring Claude onto AMD MI355X GPUs in about three days, with the model writing or porting nearly all required kernels. If that investor-presentation claim holds up, it could reduce the engineering cost of using more than one accelerator vendor; the result has not been independently verified.

## Summary
Anthropic says its frontier model Mythos helped bring Claude onto AMD MI355X GPUs in about three days, with the model writing or porting nearly all required kernels. If that investor-presentation claim holds up, it could reduce the engineering cost of using more than one accelerator vendor; the result has not been independently verified.

## What happened
The Decoder reports that Anthropic described the work in an investor presentation, with co-founder and Chief Compute Officer Tom Brown explaining the company's multi-chip strategy. Anthropic says it trains and runs Claude across four hardware families: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. It treats capacity across vendors and locations as an interchangeable pool for research, model development, customer demand and internal Claude use.

The most specific example concerns AMD GPUs. Anthropic says it used its frontier model Mythos to port Claude to AMD MI355X accelerators. The model wrote or ported nearly 100 percent of the kernels, and the model was running on AMD after about three days, with almost no human intervention. The measured workload described in the presentation is prompt processing, not a claim that every training and inference workload has been migrated or performs equally well on all four platforms.

Anthropic contrasts the three-day effort with its characterization of conventional onboarding: dozens of engineers occupied for months, and six to twelve months to bring a new accelerator to a usable performance level. The company says it has spent years building the custom software and inference stacks required for each chip architecture. Its argument is that model-assisted engineering now makes that accumulated multi-chip work more valuable by shortening a traditionally expensive part of a hardware transition.

The presentation also frames multi-vendor capacity as a way to choose hardware by price-performance and availability, including using different chips for training and inference. Anthropic says it gets more out of each compute dollar than any other frontier lab, but the article does not provide a comparable measurement for that claim. The announcement describes no price for the porting work, no public tool or process for other teams to reproduce it, and no benchmark comparing the resulting AMD deployment with Nvidia or other accelerators.

## What the archive adds
New Runtime's earlier coverage reported AMD's $1 billion investment in Meta and Meta's testing of Helios rack-scale systems ([November 1 note](https://newruntime.com/posts/amd-will-invest-1-billion-in-meta-s-a47dba/); [November 2 note](https://newruntime.com/posts/amd-and-meta-advance-strategic-partnership-to-power-066856/)). Those were AMD partnership and infrastructure developments; Anthropic's reported port is a new event, extending the discussion from chip supply to the software work needed to use another accelerator family.

## Trend signal
This is a weak, early signal for [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/): Anthropic presents model-assisted porting and its accumulated per-chip software stacks as a way to make its compute pool less dependent on one vendor. The reported test is narrow, and does not yet establish that other labs can reproduce the result or that hardware choice is broadly interchangeable.

## Why now
The Decoder places the account in Anthropic's investor presentation ahead of its planned stock-market debut. The company is making a case that its compute strategy can span multiple hardware families, at a time when the article describes accelerator supply as tight and notes Anthropic's existing AMD deployment plans.

## Connections to other events
New Runtime's earlier coverage of AMD's strategic partnership with Meta described Helios systems being installed and tested there ([November 2 note](https://newruntime.com/posts/amd-and-meta-advance-strategic-partnership-to-power-066856/)). That coverage concerned deployment infrastructure; Anthropic's account adds a distinct claim about using a model to reduce the software-porting work involved in using AMD hardware.

## Why it matters
For teams building AI products, accelerator choice is not determined by chip specifications alone. Custom kernels, inference stacks, performance tuning and the engineering time required to validate a new platform can make a theoretically cheaper or more available accelerator impractical. Anthropic's claim, if it proves transferable, points to a change in that calculation: models could help produce some of the low-level adaptation work that has historically made switching costly.

The caveat is important. A three-day port of prompt processing on one accelerator is not evidence that an arbitrary workload can be moved in three days, that quality and throughput remain acceptable, or that total cost falls after tuning and operations. Teams should treat the account as a reason to test portability, not as proof that multi-vendor infrastructure is now easy. The more useful evidence would include reproducible workload details, performance and cost comparisons, and the human review needed to trust generated kernels.

## Limits and uncertainty
The figures come from Anthropic's investor presentation and have not been independently verified. The account does not establish how the nearly 100 percent kernel figure was measured, what human review or debugging followed generation, or whether the three-day period included all production validation. It describes prompt processing on MI355X GPUs; results for other models, training workloads or serving conditions are not established. The article also reports Anthropic's broad compute-efficiency claims without comparative data that would let readers assess them.

## Reader takeaway
Anthropic is making a significant claim about Mythos helping Claude run on AMD hardware in days rather than months. For now, the key signal is that model-assisted porting may lower one barrier to multi-vendor compute—not that accelerator switching has become routine or that Anthropic's reported timeline is independently confirmed.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)