Frontier labs rethink AI pricing models as budgets tighten

The companies selling AI to corporate America have hit an awkward problem. Their customers are using the product so heavily that they are running out of money to pay for it.

According to the Financial Times, AI tools are chewing through enterprise budgets fast enough to force a rethink of how frontier labs price their models. That rethink is unfolding as OpenAI and Anthropic move toward initial public offerings.

The price cuts, by the numbers

The discounts have been steep. OpenAI cut pricing on its GPT-5.6 Luna model by 80% in late July, bringing it to $0.20/$1.20 per million tokens.

Those paired figures follow the usual industry format. The first is the cost of text you send the model, and the second is the cost of text it sends back.

Anthropic followed with its own moves. In August 2026, it made Sonnet 5’s pricing permanent at $2/$10 and reduced Opus 5.5 to $4/$20, a change that delivered nearly 40% savings through efficiency gains.

Then came a concentrated 72-hour stretch in early September 2026. During that window, Anthropic cut cache-read pricing on its Fable 5.1 model by 75%, to $0.25 per million tokens.

Cache reads are the AI equivalent of a coffee shop remembering your usual order. The model reuses context it has already processed, and customers pay less for that repeated work.

The broader market reflects the same pressure. Some analyses show blended token prices falling between 41% and 64% from mid-2026 peaks, squeezing premium pricing strategies across the sector.

Why corporate budgets cracked

Demand is the root cause. Token volume has grown 25-fold since 2023, as businesses moved AI from pilot projects into daily workflows.

Uber offers a vivid example. The company reportedly exhausted its annual AI budget in just four months.

Enterprises have responded by introducing internal spending caps and routing workloads to cheaper models. Mid-tier models now deliver about 90% of flagship performance at a fraction of the cost.

Background: a market built on falling costs

Price declines are not new in this industry. The cost of equivalent AI performance has been falling approximately 47% per quarter.

What is different now is the competitive framing. Rather than quietly passing along efficiency gains, the labs are cutting prices in visible, rapid succession, and the moves look increasingly like a contest for enterprise wallet share.

The research describes the dynamic as a pricing war. The mid-tier segment, once a middle ground between budget and premium offerings, has become the main battlefield.

What this means for the labs and their customers

For enterprises, the short-term picture is favorable. Lower prices and capable mid-tier models give procurement teams more room to expand AI usage without blowing through budgets in a single quarter.

For OpenAI and Anthropic, the stakes are sharper. A company heading to public markets wants to show durable revenue and pricing discipline, and repeated cuts of 75% and 80% complicate that story.

When mid-tier models reach about 90% of top-end performance, the premium for that last 10% has to be justified on very specific workloads. Labs will need to show that their most expensive models do things the cheaper ones cannot, or accept that the premium segment shrinks.

Key signals to watch include whether the price cuts stabilize or continue, how enterprises adjust their annual AI budgets, and how the labs explain their pricing strategy as their listings approach. The September cuts arrived within a 72-hour window, and the pace of change suggests the next round may not be far behind.