OpenAI and Anthropic slashed frontier Artificial Intelligence (AI) token prices on September 22 [1][2]. But 472 business buyers told the Tokenomics Foundation the next day that price was not the problem. Only 4% asked for cheaper rates, while 23% demanded a bill they could explain to their Chief Financial Officer (CFO) [3].

I have watched this shift all week, and that 4% figure is the part I keep circling. The two biggest AI labs raced their prices to zero on the same day, and buyers just shrugged.

Why it matters: The sticker price no longer shows the real bill. A token rate hides the true costs of enterprise AI, like the agent loops, cache misses, and human reviews that sit between a prompt and a finished job. The token price is the only line item shrinking, and it was never the expensive part.

OpenAI says better caching and inference funded its Generative Pre-trained Transformer (GPT) price cut [1]. Anthropic is also discounting cache reads because they drive “the majority of agentic and coding work costs” [2].

Price is not the problem

The big picture:

The Tokenomics Foundation — a Linux Foundation project — reports that three in four enterprises cannot prove AI outcomes to their CFO [3]. A full 39% cannot link AI spending to any clear result at all [3]. The core problem is a lack of proven Return on Investment (ROI), not a high price tag.

By the numbers:

  • 50%: OpenAI dropped GPT-6 Sol to $2/$10 and GPT-6 Luna to $0.10/$0.50 per million tokens, down from $4/$20 and $0.20/$1.20 [1].

  • 40%: Anthropic claims Opus 5.5 costs 40% less to run than Opus 5 while matching its pricier Fable 5.1 model on most work [2].

  • 60%: Anthropic cache read costs fell to $0.20 per million tokens [2].

  • 4% vs. 23%: The tiny share of buyers asking for cheaper prices versus the group begging for clear billing data [3].

Token Prices Just Halved. The CFO Still Can’t Read the Bill.Figures as stated in this article’s own numbers section (verified figures, %)OpenAI GPT-6 Sol50%([1])Anthropic cache read costs60%([2])Buyers asking4%([3])

Where this bites:

The operators closest to the story are changing how they track AI costs. The survey found that 88% of enterprises now have a named owner for AI economics [3]. My read is that this marks the first real fix, as having a dedicated owner makes a company 3.7 times more likely to show the CFO real value [3].

  • Model routing speeds up savings: Right now, 86% of enterprises use or are testing a model router [3]. Router users are four times more likely to prove value to the CFO [3]. Sending each task to the cheapest capable model ensures the price cut actually reaches the bottom line.

  • Buyers want standard bills: Only 4% of buyers asked for discounts, but 7% asked for FinOps Open Cost and Usage Specification (FOCUS) billing [3]. They wrote this open cloud-billing standard into a free-text survey answer without any prompts [3]. A shared billing format is worth more to a finance team than another 50% discount.

  • The new metric is work units: What I’d watch next is whether vendors start publishing their cost-per-resolved-work-unit. Finance teams need a clear denominator — cost per resolved ticket or merged code change — rather than cost per token. This shift turns a vague price cut into proven value.

The catch:

These big price cuts still matter to teams running a well-tracked tech stack. A 50% drop on the token bill saves real money if you already know where it goes.

The part I keep circling is how this widens the gap between winners and losers. Leaders with model routing capture every cent of this price drop.

Meanwhile, the 75% of enterprises that cannot prove AI outcomes just see the price drop without getting any closer to a solid business case [3]. I could be wrong that this forces companies to rebuild their systems, but buyers clearly want a bill they can read over a cheaper one.

The Big Shift: OpenAI and Anthropic slashed frontier AI token prices on September 22, but buyers revealed that price was never the real problem. Only 4% asked for cheaper rates, while 23% demanded a bill they can explain to finance teams.

Why It Matters:

Enterprise AI costs have moved past the raw token and into agent loops, cache misses, and human reviews. The token price is shrinking while the actual bill stays hidden, making cost-per-token a poor sign of business value.

What I’d Watch:

  • Cost-per-resolved-work-unit: Whether vendors start reporting the cost per completed task rather than the cost per token, giving finance teams a metric they can actually track.

  • FOCUS billing standardization: The open FinOps Open Cost and Usage Specification schema that 7% of buyers requested unprompted — a shared format is needed to track costs across providers.

  • Model routing adoption: 86% of enterprises are already testing or using routers to send tasks to the cheapest capable model, which captures price cuts best.

Related reading

Sources

[1] OpenAI — “Introducing GPT-6 Sol and Luna” (September 22, 2026) — https://openai.com/index/introducing-gpt-6-sol-and-luna/ [2] Anthropic — “Introducing Claude Opus 5.5” (September 22, 2026) — https://www.anthropic.com/news/claude-opus-5-5 [3] Tokenomics Foundation — “State of Tokenomics, September 2026” (September 23, 2026) — https://www.finops.org/insights/state-of-tokenomics-september-2026/