In a dramatic shift for the artificial intelligence economy, OpenAI and Anthropic have cut their flagship model API prices by as much as 80 percent. The move comes as they try to hold ground against fast-rising, low-cost Chinese rivals such as DeepSeek and Moonshot AI. What was once a race to spend whatever it took is now giving way to a tighter balance between heavy infrastructure outlays and falling inference costs that squeeze margins.
Quick Take
- The catalyst: OpenAI and Anthropic drastically cut inference costs—with OpenAI’s GPT-5.6 Luna dropping by 80%—in direct response to high-performing, hyper-cheap models from Chinese competitors like DeepSeek and Zhipu.
- The new benchmark: Average US AI-inference prices plummeted nearly 25% over a single month in late 2026, bringing GPT-4-class reasoning from roughly $20 per million tokens in 2022 to well under $1 today.
- The margin pressure: As enterprise procurement shifts to a “performance per dollar” mindset, US labs face mounting questions ahead of potential public listings about how they will recoup massive GPU infrastructure investments while defending market share.
- The under-reported angle: This deflationary spiral radically alters AI application unit economics, making complex agentic loops and massive cached-context workloads viable for developers previously priced out of continuous inference.
What happened
Over the last few months the foundational model market has slipped into a sharp price decline. OpenAI started by lowering GPT-5.6 Luna 80 percent, bringing input costs down from $1 to $0.20 per million tokens and output costs to $1.20. It trimmed GPT-5.6 Terra by 20 percent and launched GPT-6.1 Sol at half the price of the prior version.
Anthropic followed suit, locking in its Claude Sonnet 5 discounts and cutting Claude Opus 5.5 by 40 to 60 percent. Both labs also began offering steep discounts on cached input tokens, sometimes as low as 5 percent of the usual rate, which eases the cost of workloads that reuse large context windows.
The trigger was clear: competitive Chinese models from DeepSeek, Moonshot AI, Zhipu, and Z.ai. These systems deliver comparable results at far lower cost. One test showed an enterprise workload running on Zhipu GLM for roughly $544 versus $4,811 on Claude, undercutting the premium US pricing power.
Why it matters now
The market has moved from a focus on raw capability to one centered on price. For three years the story centered on building ever-stronger models and the data centers to train them. Now that top-tier costs are falling fast, labs must carry billions in infrastructure spending while revenue per unit drops.
Developers and AI startups gain breathing room. Workloads that once looked uneconomic, such as agent swarms running nonstop or large codebases held in cache, are suddenly viable. That shift lets companies either lower prices for users or keep more of the revenue themselves.
Yet the same pressure reaches cloud providers and chip makers. If inference turns into a pure race to the bottom, returns on new GPU-heavy facilities could weaken, affecting everything from capacity plans to hardware demand. At the same time, enterprises weighing Chinese models strictly on cost may draw regulatory attention to how geopolitics and open-weight platforms shape the base layers of software.
Who is most affected
The ripple effects of this price war extend far beyond the billing dashboards of software engineers, fundamentally redistributing power and risk across the AI value chain. Open-weight ecosystem hubs are thriving by facilitating easy price-comparisons, while deep-pocketed foundational labs are forced to rethink their path to profitability just as public market investors begin scrutinizing their balance sheets.
| Stakeholder | Impact | Why |
|---|---|---|
| AI App Developers | Highly Positive | Plunging token prices and massive cached-context discounts allow for complex agentic workloads and vastly improved SaaS gross margins. |
| Enterprise Buyers | Positive | Shifting from brand loyalty to “performance per dollar” frameworks gives procurement teams leverage to mix US and Chinese model APIs. |
| Tier 1 AI Labs | Negative / High Risk | Drastic price cuts heavily compress profit margins, complicating the timeline to recoup billions spent on GPU clusters and affecting IPO narratives. |
| Data Center Operators | Uncertain | While cheaper inference drives higher usage volume, compressed lab margins may slow future speculative capacity leasing and physical infrastructure build-outs. |
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent i10x analysis synthesizes current market reporting to equip CTOs, AI developers, and infrastructure investors with a clearer view of foundational model economics. We evaluated recent vendor moves, secondary pricing data, and global supply chain pressures to map the changing cost of intelligence.
i10x Perspective
The collapse of premium inference pricing shows that frontier AI is spreading faster than older defenses can contain it. If Chinese challengers and open-weight systems keep turning model intelligence into a commodity, US labs will have to compete on enterprise data lock-in and specialized agent tooling rather than token price. The key question is whether the surge in application usage can outrun the heavy capital cost of the infrastructure behind it.
