OpenAI and Anthropic in price war as Chinese AI rivals gain ground

The price war has arrived. OpenAI cut GPT-5.6 Luna by 80 percent. Anthropic launched Claude Opus 5 at half the price of its own flagship. These are not competitive adjustments. They are distress signals disguised as product launches. The Silicon Data token price index shows a 25 percent drop since mid-July. That number should terrify every investor banking on AI's pricing power.

For years the narrative held that proprietary models justified premium pricing through superior capability. Performance gaps would widen. Moats would deepen. Reality intervened. Chinese labs released models that developers can download, modify, and deploy without asking permission. The performance gap narrowed. The price gap exploded. Moonshot and DeepSeek did not need to match GPT-5.6 feature for feature. They only needed to get close enough that the premium looked absurd.

Enterprise customers noticed. DoorDash and Airbnb did not switch to Chinese models for ideology. They switched because their AI bills became unsustainable. Usage-based billing — the great hope for aligning costs with value — became a trap. Every query, every token, every millisecond of inference now carries a meter. CFOs imposed caps. Engineering teams tested alternatives. The alternatives worked.

OpenAI and Anthropic face a structural squeeze. They burn billions on compute, talent, and infrastructure to maintain a lead that shrinks by the quarter. Their investors demand trillion-dollar exit valuations. Those valuations require recurring revenue with expanding margins. The price war delivers the opposite: declining revenue per token with no floor in sight.

The closed-model moat was always thinner than its architects claimed. Performance advantages in large language models prove surprisingly transferable. Techniques diffuse. Talent migrates. Papers publish. Open weights accelerate the process. A developer in Shenzhen or Beijing can iterate on a downloaded model tonight. A developer in San Francisco waits for an API update next quarter.

National security rhetoric will not stop this. The US government can restrict chip exports. It cannot restrict mathematics. The transformer architecture is public. The training recipes are replicable. The data is scrapable. Chinese labs operate with fewer chips but greater urgency. They optimize. They distill. They release.

Anthropic's "frontier intelligence at half the price" slogan captures the contradiction. Frontier intelligence cannot simultaneously command a premium and require a 50 percent discount. The market decides which claim is real. So far the market chooses price.

OpenAI's 80 percent cut on Luna admits the same truth. The "fastest and most affordable" framing cannot hide the desperation. When the leader cuts prices that deeply, the leader acknowledges that its differentiation no longer justifies its margin.

The IPO timeline makes this dangerous. Both companies need to show hockey-stick revenue growth before public markets scrutinize their S-1 filings. Usage-based billing was supposed to deliver that growth — more usage, more revenue. Instead it delivered cost consciousness. Customers treat AI like cloud compute: a commodity to be optimized, not a premium to be cherished.

Chinese open models exploit this perfectly. They arrive without contracts, without minimums, without sales calls. They arrive as weights on Hugging Face. A finetune here, a quantization there, and suddenly the cost per million tokens drops 90 percent. The performance delta? Negligible for summarization, classification, extraction — the bread-and-butter workloads that constitute most enterprise volume.

The US labs respond with feature differentiation. Reasoning tokens. Tool use. Structured output. Multimodal native. These matter for frontier applications. They do not matter for the millions of routine tasks that pay the bills. A model that classifies support tickets at 1/10th the cost wins that workload. The reasoning model stays idle.

Investors should ask what happens when the price floor hits compute cost. Nvidia sets that floor. Neither OpenAI nor Anthropic controls it. Both buy from the same constrained supply. Their margins compress between Nvidia's pricing power and DeepSeek's free weights.

The bear case writes itself: AI model inference becomes a utility. Thin margins. High volume. Capital intensity without pricing power. The bull case requires believing that capability divergence will reopen — that the next paradigm shift (reasoning, agents, world models) restores the moat. That bet cost billions last year. It costs billions more this year. The clock ticks.

Chinese labs do not need to win the frontier. They only need to win the volume. Volume pays for the next training run. Volume attracts the talent. Volume compounds. Moonshot's Kimi 2 and DeepSeek's V3 already power production systems across Europe and Southeast Asia. Their users contribute data, feedback, and distribution. The flywheel spins.

US policymakers frame this as a contest of national champions. It is not. It is a contest between proprietary APIs and open weights. The open weights are winning the adoption curve. The proprietary APIs are winning the headline benchmark curve. History suggests adoption compounds faster than benchmarks.

OpenAI and Anthropic still possess advantages: brand trust, safety tooling, enterprise support, compliance certifications. These justify some premium. They do not justify 5x premiums. The market is recalibrating. The recalibration will be brutal for valuations built on the old multiple.

The price war is not a skirmish. It is the new equilibrium. Every quarter that Chinese models improve while US models discount, the equilibrium hardens. The labs chasing trillion-dollar IPOs are fighting the last war — performance at any cost — while the market has moved to the next one: capability per dollar.