Anthropic's IPO roadshow materials include a slide claiming that its Claude Mythos Preview model worked out in about three days how to run another model on AMD's MI355X GPUs. The same slide puts traditional adaptation work for serving inference on a new compute platform at typically six to twelve months.[1][2] The slide was reported by IT之家 (IT Home) on November 11, 2026, and carried the same day on the outlet's official Tencent News account.[1][2]

Read as written, the claim is a vendor's self-description inside an active IPO pitch, and for enterprise AI buyers nothing changes today in what they can buy or what they pay. Its relevance is the cost of switching between GPU suppliers, and the practical move right now is to anchor on Anthropic's published prices and watch the offering's pricing decisions in the coming days.[3][4]

The claim covers inference porting

The sourced comparison measures inference porting, the work of making an already trained model serve traffic on new accelerator hardware. Nothing in the report describes training adaptation, so the headline framing on the hot-search lists, which has Anthropic adapting chips for training in three days, states more than the source does.[1] The report also names the acting model as Claude Mythos Preview, which matches no released Anthropic product; it is a different name from the released flagship pair, Fable 5.2 and Mythos 5.2, announced on November 9, 2026, with Mythos access gated behind the company's verification programs.[5] Which model version performed the port, and whether it involved any shipping product, is not stated in the report.

MI355X is already committed at scale

The hardware side of the claim is concrete. MI355X is an announced AMD data-center accelerator built on the CDNA 4 architecture, with 288GB of HBM3E memory and 8TB/s of memory bandwidth, reported in March 2025 to reach the market in mid-2025.[7] Oracle has committed to it at hyperscale. Larry Ellison disclosed on Oracle's fiscal 2025 third-quarter earnings call a multi-billion-dollar agreement for a 30,000-GPU MI355X cluster. He argued that Oracle can build giant AI clusters faster and more economically than competitors, and that under hourly billing a speed advantage converts directly into a cost advantage.[7] The porting claim therefore targets a platform another hyperscaler has already ordered by the tens of thousands of units, which makes it a claim about switching cost between suppliers.

A three-day port does not clear CUDA lock-in

How much weight the number can carry is set by who is saying it and what evidence sits behind it. It appears in roadshow material produced by a company seeking up to $64 billion from its listing, and the report quotes the comparison as a company claim; no measurement accompanies it in the material as reported.[1][4] The same report records commentary arguing that CUDA's advantage may already have lost its force, and then gives the components that survive a fast port. CUDA's advantage rests on a large developer ecosystem, extensive libraries, and years of optimization, and Nvidia reinforces the position by offering universities free GPUs and training, and by tying NCCL, its library for communication between GPUs, to NVLink and InfiniBand hardware, which raises the cost of migrating to other platforms.[1] A fast port of one inference workload clears none of that ground.

The published prices are what a buyer can check

For a buyer budgeting AI workloads, the numbers that move decisions are on Anthropic's published price list, and those are checkable today. Opus 5.5, announced September 22, 2026, prices input tokens at $4 and output tokens at $20 per million, 20% less than Opus 5, with cache reads at $0.20 per million tokens, 60% less than Opus 5. Anthropic attributes the drop to serving efficiency, stating that the model requires less compute to serve and will cost 40% less than Opus 5 on typical workloads at default settings, in its own tests.[6] The flagship Fable 5.2 and Mythos 5.2 pair prices at $8 and $40 per million input and output tokens, with an optional fast mode at $16 and $80; Mythos access remains gated behind the verification programs.[5] A buyer's real cost breakdown runs through input tokens, output tokens, cache reads, and usage limits, and those components are where any future claim about cheaper serving would have to show up.[6]

The offering is the near-term event to watch

Anthropic has filed its Form S-1 with the SEC and intends to list on the Nasdaq Global Select Market under the symbol ANT, with the share count and price range undetermined at filing.[3] The Financial Times, in the copy carried by IT之家, reported that underwriters are weighing a price above the $70 to $78 per share range submitted on Tuesday, November 10. That price would lift the target valuation past the roughly $1.4 trillion in the filed documents, where the fully diluted range runs $1.26 trillion to $1.41 trillion, with a listing planned for next week.[4] Strategic partners including Google, Nvidia, and Microsoft plan to invest $26 billion at the IPO price through a concurrent private placement, and the company plans about $900 billion of compute and infrastructure spending over the coming years, the cost base that makes supplier flexibility worth arguing for in a roadshow.[4]

Publication of the slide with a measurement attached, by Anthropic or AMD, would move the three-day figure from pitch to evidence. So would a product or pricing announcement that puts AMD hardware inside a purchasable Anthropic stack. Until either happens, the figure remains a negotiating position, the published price list is the checkable baseline, and the switching-cost argument between GPU vendors stays worth tracking.