Claude Mythos ports AMD MI355X kernels in 3 days instead of a year — near-zero human intervention
🔎 Three days instead of twelve months, announced at the exact moment Anthropic is courting Wall Street
Anthropic's IPO roadshow delivered this week's most strategic announcement of the year: Claude Mythos Preview got an Anthropic model up and running on AMD Instinct MI355X GPUs in roughly three days. The same work used to take six to twelve months with the traditional engineering approach. And "nearly 100% of the kernels" — the code that lets the model talk to the silicon — were ported or written by the model itself, with "practically no human intervention," according to Tom Brown, co-founder and Chief Compute Officer of Anthropic.
The news, spotted by AI Weekly on November 17 via Tom's Hardware, lands at a carefully timed moment: Anthropic is finalizing its IPO on the Nasdaq under the ticker ANT, with a final price expected on November 18. An investor presentation is a sales exercise; every slide in it is an argument.
But reducing the story to pre-IPO marketing would be an analytical mistake. Because even stripped of its commercial emphasis, this announcement touches the industry's best-kept secret: the CUDA wall, that software ecosystem which guarantees Nvidia margins above 70% on AI chips. If switching silicon becomes a matter of days rather than years, the entire economics of AI compute gets renegotiated.
The essentials
- The fact: Claude Mythos Preview got an Anthropic model operational for prompt processing on AMD MI355X GPUs in ~3 days, versus 6 to 12 months historically (Tom's Hardware, November 2026).
- The striking figure: "almost 100% of kernels" ported or written by the model, with virtually no human intervention — an Anthropic claim, not an audited measurement.
- The actual scope: prefill for a single model, on a single GPU family. Not training, not decoding, not ops (SXF analysis).
- The strategy: Anthropic trains and serves Claude on four chip families (AWS Trainium, Google TPUs, Nvidia GPUs, AMD GPUs) managed as "a single fungible compute pool".
- The context: IPO on the Nasdaq (ticker ANT), AMD investing up to $5B and set to supply up to 2 GW of MI450s starting in H1 2027; Nvidia subscribing to $10B of shares in the concurrent private placement.
- Our take: even stripped of its emphasis, this result signals a structural erosion of the CUDA lock-in — slow, but now measured in days.
Tools and players: mapping the announcement
Before the analysis, let's put the pieces back on the chessboard.
| Tool / platform | Role in the story | Price (November 2026) | Key takeaway |
|---|---|---|---|
| Claude (Mythos) | Agentic model that ported/wrote the kernels | Usage-based via API (see anthropic.com) | Publicly accessible via Claude Fable 5 |
| AMD Instinct MI355X | Target GPU for the port | N/A, datacenter quote only | The demo focuses on prompt processing |
| AMD ROCm | Software stack rivaling CUDA | Open source | Accelerated with Claude (partnership July 2026) |
| Nvidia CUDA | Dominant ecosystem to unseat | Free, but lock-in | Nvidia's margin >70% on AI chips |
| AWS Trainium / Google TPU | Anthropic's other chip families | Cloud contracts | Integrated into the fungible compute pool |
Two important clarifications: the demonstration concerns prompt processing (prefill), not the full stack. And ROCm, while open source, remains far from CUDA's ecosystem maturity — this gap is precisely what the partnership with Claude claims to close.
What Anthropic Actually Announced
In three days, Claude Mythos Preview got an Anthropic model operational for prompt processing on AMD MI355X GPUs — work that historically took dozens of engineers six to twelve months.
The facts first. The announcement comes from the IPO roadshow investor presentation, picked apart by Tom's Hardware and then picked up by Mezha and The Decoder. In it, Tom Brown lays out the stakes: "we now train and serve Claude on all four major chip families... each chip requires its own software and inference stack, which is costly and difficult to build" (transcript published by Simon Willison on November 11, 2026).
Against that backdrop, one number: "Claude ported or wrote almost 100% of the kernels" needed to bring an Anthropic model to a useful level of prompt processing performance on the MI355X, in about three days. Historically, this work required "dozens of engineers," Brown recalls. Anthropic began working with AMD GPUs in 2026.
The detail that changes everything
The three-day figure covers prompt processing — the prefill — of a single model, on a single GPU family, up to a level of performance deemed useful. It's a critical stage of inference, the one that processes your input context. But it is neither training, nor token-by-token decoding, nor distributed communication, nor production operation.
The IPO prospectus, subject to SEC constraints, uses markedly more measured language: bringing a new chip platform into production "historically required intense engineering effort over many months" (MLQ.ai). The slide sells the dream; the prospectus sells regulatory realism. Both documents agree on the substance: porting is expensive, and something has just made it a lot less expensive.
| Item | Traditional porting | Claude Mythos porting |
|---|---|---|
| Duration (prompt processing, MI355X) | 6 to 12 months | ~3 days |
| Staff involved | "Dozens of engineers" | "Virtually no human intervention" |
| Kernels ported/written | Human work | "Almost 100%" (Anthropic's claim) |
| Independent verification | N/A | None to date |
| Scope | Full stack, multi-stage | Prefill of one model, one GPU family |
Why the CUDA wall is the real target
The technical feat matters less than its target: Nvidia's lock-in economics, which rest entirely on the cost of exiting CUDA.
For twenty years, CUDA has been the most effective defensive moat in tech. Every AI lab, every hyperscaler, every inference platform writes code that only runs on Nvidia. Switching GPU vendors means rewriting thousands of optimized kernels: months of effort, dozens of engineers, enormous industrial risk. That exit cost is what funds Nvidia's margins of over 70% on AI chips (The Decoder, November 2026).
What Anthropic's announcement says is that this exit cost is collapsing — at least for a portion of the pipeline. A port that costs twelve months and a dedicated team is a strategic decision, weighed at the executive committee level. A port that costs three days is routine operations. And when switching vendors becomes routine, the commercial balance of power tips.
That's the whole point of the partnership signed in July 2026: up to 2 GW of MI450 GPUs deployed starting in the first half of 2027, and a ROCm accelerated with Claude's help (MadRobot, November 2026). ROCm has never lacked ambition; it lacked cadence. Compressing the porting cycle from twelve months to three days is changing the cadence.
Note the irony of the lineup: Nvidia subscribed to $10 billion worth of shares in the concurrent private placement, versus $2.5 billion for AMD (MLQ.ai, November 2026). The incumbent is funding a competitor's entry — proof that even Nvidia no longer considers its wall eternal. And diversification is a structural trend that goes beyond silicon: on the model side too, the balance is shifting, as our analysis of Qwen4-Max, Alibaba's new flagship shows.
Four chip families, a single "fungible" pool
Anthropic's hardware diversification is not a technical precaution: it's a negotiating weapon.
The most under-discussed detail of the announcement is not the three days. It's Tom Brown's phrasing: the four chip families are managed as "a single fungible compute pool". One unit of compute is worth another, regardless of the logo etched into the silicon.
The implications are economic before they are technical. A buyer able to shift their workload between Trainium, TPU, Nvidia and AMD doesn't negotiate like a buyer captive to a single ecosystem — not on price, not on allocations, not on timelines. In a market where capacity is booked in gigawatts and energy is worth more than silicon, this flexibility is measured in billions.
And the loop closes in on itself: the cheaper the porting, the more credible the pool; the more credible the pool, the more leverage Anthropic has vis-à-vis its suppliers; the stronger the leverage, the more worthwhile it becomes to automate the porting. Mythos is not an R&D gadget. It's the centerpiece of this machinery.
"Almost 100% of kernels": what the number doesn't say
The figure is spectacular; its scope is narrow, and nobody outside Anthropic has verified it.
The most clear-eyed analysis of the matter comes from SXF, which lays out the guardrails: the result covers the prefill of one model on MI355X, not a full migration — nor training, nor decoding, nor distributed communication, nor operations. "Almost 100% of kernels" is Tom Brown's narrative, not a measured percentage of the entire inference stack codebase. No reproducible benchmark, no independent audit published to date.
The Decoder also reminds us: the figures come from an investor presentation and have not been independently verified. That's the legal framework of any pre-IPO communication — a pitch, not a peer-reviewed publication.
Should the announcement be dismissed because of this? No, and that's precisely the journalistic exercise here. Even reduced to its most conservative reading, the dynamic remains massive: weeks of human work compressed into days of machine time, on a task considered among the most demanding in software engineering. The right stance is neither credulity nor denial, but waiting for three pieces of evidence: a reproducible benchmark, an extension of the result to decoding and training, and a reproduction by a third party.
Mythos, showcase for agentic AI — and the future of code
Beyond silicon, the announcement is above all a sales pitch for Mythos, the family of agentic models that Anthropic is now pushing out to the public.
Mythos is no longer a lab secret since Claude Fable 5 made it available to the general public. And its capacity for autonomous action has already proven itself elsewhere: the security agent Glasswing identified more than 10,000 critical vulnerabilities in a single month — a volume that would keep a security team busy for years.
Writing GPU kernels is a whole notch above in terms of difficulty: systems code, where every mistake costs hours of debugging and every microsecond is counted in dollars at cluster scale. If Mythos holds up at this level on a real industrial workload, the sales pitch writes itself: the best model for coding is no longer the one that writes your application, but the one that writes the software that runs your models.
For context: our round-up of the best LLMs for coding ranks GPT-5.5 (98.2) ahead of Gemini 3 Pro Deep Think (95.4) and Claude Opus 4.7 (Adaptive) (94.3). The frontier is now shifting toward long-horizon agentic work — exactly the territory of the models we vet in our selection of the best LLMs for agents.
And the trajectory extends beyond software. In the physical world, the same logic of autonomous agents is taking hold: Figure Helix-02 has shown humanoid robots completing full eight-hour shifts in factories, without any human intervention. Everywhere, the same sea change: entrusting autonomous systems with tasks once deemed too critical to delegate.
What This Means for Developers
In six months, your stack won't change an iota. In two years, expect three very real effects.
First effect: ROCm will progress at an unprecedented speed. AMD has announced it is accelerating its stack with Claude, and the MI355X demonstration serves as proof of concept. What is built for Instinct datacenter will eventually trickle down to consumer Radeon GPUs. If you run models locally, this is good news: the classic pain point of the AMD GPU—an ecosystem designed for CUDA first—could be mitigated. Our guide to installing a local LLM and our selection of the best local LLMs will be updated accordingly.
Second effect: pressure on computing prices. When the biggest buyers can switch providers in a matter of days—and Anthropic orders up to 2 GW of MI450—monopoly power erodes. Inference will not become free, but the downward slope of prices will accelerate. You will benefit, indirectly, with every API bill.
Third effect: choice. Between Claude (Opus 4.6 and 4.7, Sonnet 4.6), GPT-5.4 and GPT-5.5, Gemini 3.1 Pro, DeepSeek V4 Pro, Grok 4.1, or GLM-5.1, the difference is now made on usage details rather than raw capacity. Our comparison Claude vs ChatGPT remains the most read entry on the site—and for good reason: these are the two ecosystems that most of you are putting into production.
And to test all of this without a cluster budget: an application that consumes these APIs runs on a simple VPS. Hostinger remains our default recommendation for getting started—a few euros per month is enough to connect your first agents to Claude or GPT (rates observed in November 2026, check on hostinger.com).
The dates to watch
Three dates will shape the coming months — each one capable of validating or refuting the announcement.
- November 18, 2026 — IPO pricing: listing expected on the Nasdaq under the ticker ANT. Our breakdown of Anthropic's official S-1 details what the company actually disclosed to the SEC — with far more cautious wording than the roadshow slides.
- First half of 2027 — MI450 ramp-up: up to 2 GW announced. The real industrial test will come then: will the porting extend beyond prefill?
- 2027 — the technical moment of truth: decoding, training, distributed communication. If Mythos pulls it off with the same human-time ratio, the conversation changes dimension.
And one discreet but decisive variable: the emergence of an independent audit or a reproducible benchmark. As long as neither exists, "three days" remains a roadshow narrative — remarkably well-crafted, but a narrative.
❌ Common Reading Mistakes
Mistake 1: believing that "AI has replaced kernel engineers"
What's wrong: the port concerns the prompt processing of a single model on a single GPU family. The rest of the effort — training, decoding, ops — has historically taken dozens of engineers, and it remains to be done on those fronts.
The fix: always read the scope before the number. "3 days" without "prefill of one model on MI355X" doesn't mean the same thing.
Mistake 2: taking "3 days" for an audited benchmark
What's wrong: the source is an investor presentation in the middle of an IPO. The Decoder notes this explicitly: no independent verification. The prospectus itself speaks of "an intense engineering effort over several months."
The fix: wait for a reproducible benchmark or an audit before citing the figure as established fact.
Mistake 3: concluding from this that CUDA is dead
What's wrong: CUDA remains the standard ecosystem, Nvidia maintains margins above 70% and has subscribed $10B to Anthropic's IPO. A moat that is eroding is not a moat that has been filled.
The fix: track the cadence of ports, not the headlines. It's the trajectory over 8 to 12 quarters that counts.
Mistake 4: ignoring the second-order effect on prices
What's wrong: many will see this news as nothing more than a lab anecdote, with no impact on their own costs.
The fix: GPU compute is the largest line item in AI spending. Any technology that lowers the cost of switching providers eventually shows up in your API and cloud bills.
❓ Frequently Asked Questions
What exactly is Claude Mythos?
Anthropic's next-generation family of agentic models. It was Mythos Preview that performed the AMD port announced at the roadshow. The general public can access it via Claude Fable 5. Anthropic positions it as its long-horizon agent system: code, security with Glasswing and its 10,000+ vulnerabilities discovered in a single month, and now GPU kernels.
Is the "3 days" figure reliable?
It comes from the IPO roadshow investor presentation and has not been independently verified (The Decoder, November 2026). It covers only prompt processing of one model on the MI355X, at a performance level deemed useful. The IPO prospectus, subject to legal constraints, remains more cautious. Reliable as a directional indication, not as a certified measurement.
Which AMD GPUs are involved?
The demonstration involves the Instinct MI355X. The July 2026 partnership calls for deploying up to 2 GW of MI450s starting in the first half of 2027. Anthropic began working with AMD GPUs in 2026 and now manages them as one of the four families in its fungible compute pool, alongside Trainium, TPU, and Nvidia.
Is Nvidia really threatened?
In the short term, no: CUDA remains the standard and margins exceed 70%. In the medium term, the risk is the shrinking cost of switching: if porting to ROCm takes days instead of a year, Nvidia's price argument weakens. The fact that Nvidia is subscribing $10B to Anthropic's IPO says a lot about its own reading of the risk.
What is ROCm, in simple terms?
AMD's open source software stack, the functional equivalent of CUDA: it's what allows AI applications to leverage AMD GPUs. Its historical weakness lies in ecosystem lag and the slowness of ports. This is precisely what the partnership with Claude — and the MI355X demonstration — aim to accelerate.
When will we know what happens next?
The final IPO price was expected on November 18, 2026, with listing under the ticker ANT. The MI450 deployment starts in H1 2027. Above all, watch for the extension of the port to decoding and training: each step validated independently will strengthen — or invalidate — the three-day narrative.
✅ Conclusion
By compressing a twelve-month GPU port into three days — even on a narrow scope and without an independent audit — Claude Mythos has just shown that the CUDA wall is cracking from here, and that the next battle in AI compute will be fought in software written by AIs.
Stay focused on what directly concerns you: the choice of model. Our monthly comparison of the best LLMs is updated as new releases come out — that's where your performance and your bill are decided, whatever the silicon under the hood.