# AI Inference Prices Split Roughly 100-to-1 as OpenAI Pushes a New Value Metric

Chamath Palihapitiya pegs a million tokens at about 56 dollars from Anthropic and 50 cents from Chinese models, while OpenAI's finance chief tells buyers to stop counting tokens.

- Published: 2026-07-19T05:30:22.062Z
- Canonical: https://polylog.news/ai/2026-07-19/ai-inference-prices-split-roughly-100-to-1-as-openai-pushes
- Publisher: Polylog (AI desk)
- Section: markets
- Sources: [Polylog editors](https://polylog.news), [CNBC](https://www.cnbc.com/2026/07/14/chamath-palihapitiya-ai-tokenmaxxing.html), [Axios](https://www.axios.com/2026/07/17/openai-ai-costs-roi-metrics)

The most revealing view of AI economics this week came not from a benchmark but from a pricing chart. On the All-In podcast and in [CNBC commentary](https://www.cnbc.com/2026/07/14/chamath-palihapitiya-ai-tokenmaxxing.html), investor Chamath Palihapitiya described a million tokens as a "barrel of intelligence" and listed the range: roughly 56 dollars from Anthropic's best model, about 26 dollars from OpenAI, 1.50 from Meta, 1 dollar from xAI and Google, and 50 cents from leading Chinese models, as [surfaced on AI Post](https://t.me/aipost/7561). He called it the steepest commodity price decline he has seen, a claim worth attributing to him rather than treating as established fact.

  Two of his supporting claims are assertions rather than verified figures. He said token costs are falling fast enough to roughly halve on a short cycle, and that measured enterprise productivity gains sit near 5 percent. Neither has been independently reproduced, and both support his warning to companies locked into expensive vendor contracts.

  On the same days, OpenAI Chief Financial Officer Sarah Friar proposed a different measure. In an [Axios interview](https://www.axios.com/2026/07/17/openai-ai-costs-roi-metrics) and an OpenAI post titled "a scorecard for the AI age," she argued that buyers should track "useful intelligence per dollar" and ask whether AI completes work that matters and what each successful task actually costs, including retries and human review. The argument is coherent, and it also moves the comparison away from the per-token price, the measure on which American frontier labs are most exposed to lower prices from Chinese suppliers and from xAI.

## What this means

The channel here is cost and distribution. If intelligence behaves like a commodity, the closed US labs that charge premium per-token rates lose pricing power to open-weight Chinese suppliers and to xAI and Google. OpenAI and Anthropic therefore have a direct incentive to reframe value around completed tasks rather than raw tokens. The enterprises most exposed are those that signed early fixed-price contracts and now try to pass through a falling input cost as if it were stable.

## What to watch

- Whether large enterprises actually adopt a task-based cost metric in procurement, which would signal that buyers accept the labs' framing over raw token comparisons.
- The lowest level of Chinese model pricing, because a price that keeps dropping while capability holds would widen the gap the US labs are trying to downplay.