For three years, every frontier AI model release was cheaper than the one before it. On April 24, 2026, OpenAI launched GPT-5.5 at $5 per million input tokens — exactly double the $2.50 it had charged for GPT-5.4 six weeks earlier. The same day, on the other side of the planet, DeepSeek released V4 Pro at $1.74 per million input tokens. One vendor doubled. The other undercut. They had been competing in the same market for a year.

They are not in the same market anymore.

What the Race Was For

The frontier AI market that existed from late 2023 through early 2026 was structured as a race to zero. Capability differentiation was real but narrow at the top of the curve. The marginal customer — a developer evaluating providers, an enterprise comparing total inference cost, an aggregator routing queries to whichever model was cheapest at the relevant capability tier — chose on price. Every major lab understood this and behaved accordingly.

The trajectory was monotonic. GPT-4 launched in March 2023 at $30 per million input tokens. GPT-4 Turbo dropped that to $10 in November 2023. GPT-4o landed at $5 in May 2024. GPT-4o-mini opened a new tier at $0.15. DeepSeek's January 2025 release compressed the floor again. Anthropic priced Claude 3 Opus, Sonnet, and Haiku across the new compressed band. Alibaba halved Qwen3-Max in November 2025. By March 2026, the GPT-5.4 price was $2.50 input. Every quarter for thirty-six months, the floor moved down or stayed flat. It never moved up.

The structural assumption was that any provider that raised prices would lose share to whichever provider had not. The assumption was empirically correct for thirty-six months. It is no longer correct.

The Phases

Each phase looked like an ending. Each was a transition.

The first phase looked like an ending: every release a new low, the price floor approaching zero, the market converging toward inference-as-utility. The second phase looked like an ending: the floor stabilized through 2025, vendors competed on capability tiers and free-tier limits, the structure of the race seemed to settle. The third phase is what an ending stops looking like — same-day releases at a 2.87x input ratio and an 8.6x output ratio between two vendors that, six weeks earlier, had been priced within a factor of two.

What Changed Underneath

Three structural forces did the work, and none of them were visible in the price floor on the day GPT-5.4 was the marginal listing.

The first was supply. Microsoft and Anthropic absorbed enough Nvidia GPU capacity over the last six months that AI startups now report difficulty accessing frontier compute. When the buyer of marginal compute is a hyperscaler with a strategic priority rather than a startup with a budget, the supplier prices to the strategic priority. The pricing power that DeepSeek-class competition had previously transmitted to frontier vendors collapsed when the marginal customer at the top of the curve stopped being price-sensitive.

The second was capability separation. DeepSeek's own release blog for V4 Pro stated that the model "trails the performance of state-of-the-art models by about 3 to 6 months." That is a remarkable concession from a Chinese frontier lab — and a structural one. The race-to-zero pricing assumed parity at the top. The new pricing assumes differentiation at the top. DeepSeek is not pricing as a substitute for GPT-5.5. It is pricing as a substitute for GPT-5.4 — three to six months delayed — at a fraction of the cost. That is a different product, and DeepSeek is selling it as one.

The third was geography. The frontier AI customer base is not actually a single global market. It is a US-and-allies block (where data residency, export-control regimes, and procurement rules favor US vendors), a China-and-aligned block (where the same forces operate in reverse), and an emerging sovereign block where neither incumbent is acceptable. Cohere and Aleph Alpha announced a $20 billion merger this week explicitly to build a third pole — Canadian-German "sovereign AI" with both governments' support. Three regions, three different sets of buyers, three different procurement constraints. A single price could only persist while those constraints were ignorable. They are not ignorable anymore.

The Reversal

The frontier AI market was designed, implicitly, as a single global market for a near-substitutable commodity. The pricing reflected that design: monotonically declining, internally competitive, regional differences absorbed into the price floor. By April 24, the design no longer matched what the market was. The buyers had segmented; the suppliers had segmented; the prices had not yet caught up to either.

Now they have. GPT-5.5 at $5 input is a price for a buyer with a US enterprise contract, an export-controlled inference workload, and a preference for the leading-edge model over the trailing-edge one. DeepSeek V4 Pro at $1.74 is a price for a buyer with a different geographic posture, a tolerance for three-to-six-month capability lag, and an inference budget that scales with use rather than with strategic positioning. The Cohere-Aleph Alpha merger is a price still to be discovered, for a buyer who refuses both of the above. Three product-market fits. Three prices. Different markets.

This is the same dynamic that ended the cloud-computing race-to-zero in the mid-2010s. Compute was supposed to become commodity infrastructure with a single price, and for several years it did. Then data sovereignty rules and regional tax structures and customer-specific procurement preferences split the market, and the global price became three or four regional prices, each of which had its own floor. AI inference is running the same play five years faster.

Counter-Evidence

The strongest argument against calling this a regime change is that GPT-5.5 is a different product than GPT-5.4. Higher capabilities, better agentic performance, more complex reasoning support — the doubling of price could be explained as a tier shift, not a reversal of the trend. OpenAI itself has framed it that way: GPT-5.4 remains available; GPT-5.5 is the new top.

That argument is partially correct and structurally beside the point. The race-to-zero hypothesis was that the new top would be priced near the old top, with the old top falling further. What happened instead was that the new top was priced at twice the old top, while the old top stayed flat, while a competitor's new top was priced below where the old top had been. Three vendor-product configurations released within forty-five days produced three completely different price points for substantially overlapping capability. That is not a tier shift. That is a market splitting.

The Crystallization

The race to zero ended where every commodity race ends: when the market remembers that buyers are not interchangeable.

The market designed for one race produced two.