The industry spent 12–15 September debating whether to slow frontier AI from the supply side. Data published three days earlier showed buyers had already started doing it from the demand side — for entirely ordinary commercial reasons.
The Collision of Dates
March 2026 — Price Peak
Effective price per million tokens reaches $1.15 — the 2026 peak for the blended effective price Ramp reports.
August 2026 — Token Share Peak
Frontier models (Opus, Fable, Sol) peak at 53% of enterprise tokens purchased on the platform.
Frontier token share has already fallen to 45%. Effective price is $0.68/MTok — down 41% from March. Standard models (GPT-5.6 Terra, Claude Sonnet series) are carrying the volume.
12–15 September 2026 — Supply-Side Pacing Debate
The industry argues over release cadences, embedded evaluator commitments, and antitrust waivers permitting rivals to coordinate — producer-side mechanisms, every one of them.
Ramp AI Index — September 2026
45%
Frontier token share (down from 53% peak in August)
$0.68
Effective price/MTok (down 41% from March peak of $1.15)
The Ramp AI Index for September 2026, published on 9 September, showed that frontier models — which Ramp describes as models like Opus, Fable and Sol — had fallen to 45% of the tokens businesses were purchasing on Ramp’s platform, down from a peak of 53% in August — though still up from where it stood at the start of July, so this is a pullback from a peak rather than a one-way decline. The work migrating away from the frontier is going somewhere specific: cheaper standard models, with GPT-5.6 Terra and Claude’s Sonnet series carrying the incremental volume. The blended effective price is now $0.68 per million tokens, 41% below the March 2026 peak of $1.15. Two things are driving that at once, and Ramp names the more direct one first: OpenAI and Anthropic have both announced a series of price cuts over the last month. A shift in mix toward cheaper models pushes in the same direction. Both are at work, and the data here does not separate them.
Ramp’s lead economist, Ara Kharazian, described the mechanism driving this directly: “We’ve heard from businesses who are imposing company-wide defaults that reduce usage of frontier models, saying standard models are still highly performant and also more cost effective.” What Ramp describes is a company-wide default rather than a query-by-query judgement, which locates the decision with whoever configures defaults rather than with individual users.
One methodological point is non-negotiable before reading anything else into this data. Ramp measures businesses on its own platform — a large sample of US corporate card spend, not the whole market. The company notes that its top-1% spending estimates are more volatile than median figures and subject to revision. On business penetration — the share of businesses paying anything at all to a given provider — Anthropic stands at 43.8% against OpenAI’s 39.8%, with Anthropic growing faster over the month (+0.34 percentage points versus +0.09). Ramp’s own characterisation is that OpenAI underperformed overall AI adoption on this metric. That figure says nothing about how much each business spends, which models they use, or which provider leads at the frontier tier specifically. The index publishes no per-model spend-share percentages, and no such claims are made here.
The key insight: The question everyone spent last week debating — whether anyone can be made to use less frontier capability — was being answered in the background by people who were not part of the conversation and were not trying to answer it. Enterprises reached a superficially similar outcome to supply-side pacing by a completely different route: they sent a growing share of their tokens to the tier below it.
The crossover happened before the pacing argument started. Buyers did not slow the frontier down — they declined to buy it, which is a different thing with a similar surface and a completely different motive.
The Structural Read
The pacing argument that dominated 12–15 September was entirely about producers. Every mechanism on the table — release cadence agreements, embedded evaluator commitments, antitrust waivers permitting rivals to coordinate — acts on the labs building frontier models. The Ramp data describes buyers reaching a superficially similar surface by an entirely different route: not slowing what gets built, but sending a falling share of their tokens to it.
The buyer’s version has properties the producer’s version conspicuously lacks. A company-wide default routing work to a cheaper model requires nobody’s permission, involves no agreement with competitors, survives no negotiation, and takes effect the moment it is configured. It is a unilateral procurement decision, and it scales silently across every query the moment it is set.
It is worth being exact about what this resemblance does and does not mean, because the similarity is superficial in a structurally important way. Demand-side substitution reduces frontier usage. It does not reduce frontier capability. And it is motivated entirely by cost — not by the catastrophic-risk concerns the pacing argument is actually about. These are not substitutes for one another in any meaningful sense. But they compete for the same finite attention in a policy debate, and only one of them is already happening at scale.
Business Engineer Framework — Harness Theory
The Harnessers Are Already Winning on Price
Companies that harness AI rather than build it make choices the builders cannot control. The enterprise default-routing decision is a pure Harness move: take the capability that already clears the bar, configure it into the stack, and let the billing speak. The frontier labs built something good enough that the layer below it became good enough too — and a growing share of the tokens now goes there.
The second structural read is that “good enough” is always a margin event, and this is a clean example of how it arrives. Capability reaches the market at a premium. It diffuses. The tier below it catches up on the tasks that actually constitute the work queue. Buyers discover that the cheap option clears their real bar — and then, critically, they encode that discovery as a default rather than leaving it as a per-query judgment. The businesses Ramp spoke to are not saying frontier models are poor. They are saying standard models are already sufficient for the work in front of them. That is a different and more consequential claim.
Once the judgment is encoded as a company-wide default, the economics of the decision invert. Defaults are rarely revisited. The burden of proof shifts: someone must now justify buying the expensive option rather than justify switching away from it. That inversion is what could make the effect durable rather than transient: a price cut can be reversed, whereas a default that has been configured and forgotten is much harder to unwind.
Ara Kharazian — Lead Economist, Ramp
“We’ve heard from businesses who are imposing company-wide defaults that reduce usage of frontier models, saying standard models are still highly performant and also more cost effective.”
Three Implications
DEFAULTS AS STICKY GOVERNANCE
Company-wide routing defaults are a form of private governance with no appeals process and no external veto. Once a procurement or platform team encodes “use standard models by default,” that decision persists until someone actively dismantles it — and the burden falls on the expensive option. The labs cannot negotiate their way back into the default; they have to out-perform it on a cost-adjusted basis. That is a structurally different competitive problem than winning a benchmark.
MIX SHIFT VS. REPRICING — A DISTINCTION THAT MATTERS
The 41% fall in effective price per million tokens has two drivers running together: announced price cuts from both OpenAI and Anthropic over the last month, and a shift in mix as buyers choose cheaper models more often. Ramp reports the cuts alongside the decline and the data does not separate the two effects. The distinction still matters for where the revenue impact lands, because a mix shift hits the frontier tier, precisely where the labs have invested most and priced highest. The cheap tier takes a growing share of a still-growing total, while the frontier tier’s share falls. Note that a falling share is not the same as falling volume, and Ramp reports total token volume still rising; no margin data is published either way.
THE UNCOMFORTABLE IMPLICATION FOR THE PACING ARGUMENT
If frontier usage is already declining as a share of enterprise tokens for ordinary commercial reasons, then the commercial centre of gravity is shifting toward the tier below the frontier at precisely the moment the case for a slower frontier is being made. A laboratory that agrees to slow down still has to be paid, and a growing share of the work is going to a tier where frontier differentiation counts for least. Note the limits of that claim: share is not volume, Ramp reports total token volume still growing, and the index publishes nothing about revenue or margins for anyone. This is a tension, not a prediction. One month of mix shift on a single platform’s data is not a trend and may reverse. No forecast of revenue, market share, prices, or harm to any company is made here — but the tension is real and worth holding clearly.
Business Engineer Framework
Harness Theory — Where the Value Actually Accrues
The Ramp data is a Harness Theory case study in real time. Enterprises are not building the frontier — they are routing around it when it stops clearing the cost bar. Understanding which layer of the AI stack captures durable value as capability diffuses downward is the central strategic question for any operator or analyst working in this space. The Map of AI maps exactly that structure across 200+ companies and 9 layers.
The Ramp AI Index publishes no spend-share percentages for individual models. Figures circulating for particular frontier models could not be verified against the index and appear nowhere in this article; nothing here asserts that any one laboratory’s frontier model leads or trails another’s on spend. The figures of 43.8% for Anthropic and 39.8% for OpenAI measure business penetration — the share of US businesses on Ramp’s platform paying anything at all for subscriptions or tokens. They are not spend share, not token share, and say nothing about which company leads at the frontier tier. On that published measure Anthropic leads and grew faster over the month, and Ramp’s own characterisation is that OpenAI underperformed overall AI adoption. Ramp measures spending by businesses on its own platform. That is a large sample of US corporate card spend rather than a survey of the whole market, and Ramp notes its top-1% spending estimates are more volatile than median figures and subject to change. Demand-side substitution toward cheaper models reduces frontier usage; it does not reduce frontier capability, and it is motivated by cost rather than by the catastrophic-risk concerns that animate the pacing debate. The two are not substitutes for one another and nothing here suggests otherwise. The fall in effective price per million tokens has more than one driver. Ramp reports that OpenAI and Anthropic both announced a series of price cuts over the last month, alongside a shift in mix toward cheaper models; the data does not separate the two effects and nothing here attributes the decline to either alone. The frontier share of 45% is down from August’s 53% peak but up from the start of July, so it should be read as a pullback from a peak rather than a one-way decline. One month of movement on a single platform’s data is not a trend and may reverse. No forecast is offered for revenue, market share, prices, or whether any company is helped or harmed, and no model is characterised as better or worse than another. OpenAI, Anthropic and Ramp are private companies. This is businessanalysis, not investment advice, no view is expressed on any security, and no recommendation is made.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
