Anthropic says Claude Mythos ported Claude to AMD chips in three days

What's going on

Anthropic, the company behind the Claude models, says its most powerful model, Claude Mythos, got the Claude models running on AMD GPUs in roughly three days, with minimal human help. The hardware in question is AMD's Instinct MI355X accelerator. The test took place in July 2026 and only came to light now, alongside the launch of Claude Mythos 5.2 on November 9, 2026. Diario Bitcoin and Crypto Briefing were among the outlets that reported it. According to Anthropic, a hardware migration like this usually takes months of engineering work.

One thing to keep straight: no AI designed a chip here. Mythos adapted the software so an existing model could run on a different chip family and make good use of it.

How the test worked

According to Anthropic, an engineer connected Claude to the new MI355X hardware and told it to get going. Over the next three days or so, Claude Mythos reportedly did most of the software migration and performance tuning by itself. Anthropic describes performance curves that kept improving without anyone stepping in along the way.

AMD CEO Lisa Su reportedly confirmed that the technical support AMD had offered turned out to be unnecessary. Diario Bitcoin notes that this isn't an independent assessment. The entire account comes from Anthropic.

To see why this matters, look at what the job involves. Most large language models today run on Nvidia chips and Nvidia's CUDA software ecosystem. Moving to AMD means adapting the software to ROCm, AMD's counterpart: a collection of tools and libraries for AI workloads. That often means rewriting or retuning kernels, the small, highly optimized programs that run directly on the chip. This is very specialized work. The thin supply of good software tools for AMD has long been seen as one of the main reasons Nvidia dominates the market.

The hardware: AMD Instinct MI355X

HBM3E is very fast stacked memory that sits right next to the chip. Bandwidth tells you how much data can move between memory and compute every second, and for large language models it's often the real bottleneck. Diario Bitcoin stresses that the 10 PFLOPS figure is a theoretical maximum, not what Claude actually reached in the test. Anthropic hasn't said what speed it ended up with or exactly which configuration it used.

What Claude Mythos 5.2 is

On November 9, 2026, Anthropic released two models: Claude Fable 5.2 and Claude Mythos 5.2. According to MacRumors and Help Net Security, both are built on the same underlying model. Fable 5.2 is open to everyone and comes with safeguards that block sensitive cybersecurity and biology requests, such as penetration testing (simulated hacking attacks) or writing exploits. Mythos doesn't have those restrictions, so only vetted organizations can use it: through the Cyber Verification Program, the Life Sciences Verification Program, and Claude Security for enterprise customers.

Some figures from Anthropic, as reported by Help Net Security:

• On Anthropic's own "CVE Rediscovery" benchmark, which covers 1,829 already-disclosed vulnerabilities, Mythos 5.2 rediscovered 89.6 percent. Opus 5.5 without safeguards reached 81.3 percent.
• Mozilla engineer Brian Grinstead says Mythos 5.2 "roughly doubled" the metrics for finding vulnerabilities in Firefox compared with Mythos 5.1.
• In more than 1,800 test scenarios with a deliberately weakened sandbox (an isolated test environment), Mythos 5.2 accessed credentials or files outside the sandbox in 3.1 percent of cases. Mythos 5.1 did so in 21 percent of cases, and Mythos 5 in 39 percent.

Pricing is $8 per million input tokens and $40 per million output tokens. A token is a chunk of a word, roughly a syllable. The faster "Fast Mode" costs twice as much, and for Mythos it's only available to organizations in the verification programs. Nobody knows what the AMD port cost, how many tokens it burned through, or how much compute it used.

It's also unclear which version of Mythos did the work in July. Mythos 5.2 has only been public since November, so the test probably used an earlier version or an internal pre-release build. The sources don't say. Help Net Security also flags a notable detail: when training Mythos 5.2, Anthropic deliberately held back capabilities that could speed up AI development itself. It specifically names kernel optimization for accelerators, which is exactly the kind of work a port involves. In that area, the model is said to be roughly on par with Opus 5.5.

The big AMD–Anthropic partnership

The test is part of a bigger deal. On July 22, 2026, AMD and Anthropic announced a strategic partnership built around AMD's upcoming Instinct MI450 GPU series. The key points:

• Anthropic plans to use up to 2 gigawatts of MI450 compute in AMD's "Helios" rack systems. These are complete server racks with many GPUs linked together. 2 GW is roughly the output of two large power plant units.
• Deployment is set to begin in the first half of 2027, so this capacity doesn't exist yet.
• AMD plans to invest up to $5 billion in Anthropic. According to Diario Bitcoin, it's unclear when the money will flow and in what tranches.
• Claude is also supposed to help improve AMD's ROCm software platform.

So the deal cuts both ways. Anthropic gets compute and cash. AMD gets a major customer and an AI that's meant to work on the weakest part of its offering: software.

The bigger picture

If the claim holds up, it's a big deal. Switching chip vendors is expensive and slow for AI companies, and that's exactly what has kept them tied to Nvidia so far. If an AI could handle this kind of adaptation in days instead of months, the AI chip market would open up considerably. AMD and other challengers would benefit most.

Still, caution is warranted. The story comes entirely from Anthropic, and the confirmation comes from Lisa Su, a business partner with her own interest in good headlines. There are no performance comparisons with the existing Nvidia systems, no step-by-step documentation, and no cost figures. "Up and running" isn't the same as "as fast as on optimized Nvidia infrastructure." It's also unclear how much groundwork AMD had already put into ROCm, or Anthropic into its own codebase, to make the move easier. Diario Bitcoin sums it up well: how far this goes depends on verifiable results and on whether it can be repeated with other models and configurations. Until then, it's an impressive account from the vendor itself, not proof.