The token got cheaper. The task did not.

I.A price collapse nobody's invoice noticed

Silicon Data's LLM Token Expenditure Index — a daily, usage-weighted blend of what the market actually pays per million tokens across more than 200 open and 120 proprietary models, published on Bloomberg under the ticker SDLLMTK — printed $0.97 on 3 September 2026. On 31 May it was $2.04. In late July it was $1.45, and by 6–8 August it had reached $1.16–$1.18, a level Jefferies flagged as the year's low in a note reported on 20 August. Taking the index at both ends gives a fall of about 52% in fourteen weeks (our arithmetic, from the two published values).

That is a genuinely large move in a genuinely short window, and it is the kind of number that gets read as a verdict on the whole AI trade. It should not be. The index measures the average price of a token bought, not the price of any particular token, and most of its decline is composition, not repricing.

II.Two indices, opposite directions

Run the same question through a different measurement and the answer inverts. BenchLM's Token Price Index, which tracks the median list price of models at each capability tier rather than weighting by usage, had the frontier tier at 16 in September 2026 against a March 2023 base of 100 — but up 72% year on year, with a median blended cost of $6.00 per million tokens across 21 frontier models. The budget tier was up 36.8% on the year; only the mid tier fell, by 7.7%.

Both readings are correct. Frontier intelligence is 84% cheaper than it was in early 2023 and meaningfully more expensive than it was twelve months ago. The clearest single illustration came at the end of July: aggregators tracking OpenAI's published rates record GPT-5.6 Luna cut by 80% and Terra by 20% on 30 July, while the flagship Sol launched a day earlier at $5 input and $30 output per million — roughly four times GPT-5's $1.25 and $10, and the first increase in that product line's flagship price since March 2023. One vendor, one week, deflation and inflation in the same price list.

III.Who actually paid less

The honest technical read is that the deflation is real but narrowly distributed. YipitData's June 2026 analysis, built on roughly 150 trillion tokens a month — near two quadrillion annualised — found effective pricing across API services down only 6% year-to-date in 2026, against a 39% fall in the second half of 2025. Split by customer, the divergence is stark: the developer market visible through OpenRouter fell about 15% through March, while enterprise consumption via cloud and Bedrock showed no meaningful change in effective price at all. Developers routed down-tier toward Kimi K2.5 and DeepSeek V4 Flash; enterprises routed up, toward the expensive models, despite cheaper substitutes being available.

Two mechanical caveats compound this. Prompt caching now cuts input costs by roughly 90% without any change in list price, so measured "prices" move for reasons unrelated to vendor pricing decisions. And index composition shifts as models enter and exit on sustained relevance — meaning a large SDLLMTK move can record the market's substitution behaviour rather than any seller's decision. Silicon Data itself has been explicit that the series is an expenditure-weighted average, not a price list.

IV.Where the savings went: token intensity ate them

The demand side is what converts a price collapse into a bigger bill. Gartner puts agentic workflows at five to thirty times the tokens per task of a simple chatbot query; McKinsey's 2026 work goes further, putting multi-turn agentic workloads at up to a thousand times standard chat, with the final visible output accounting for only 5–15% of tokens consumed. EY's estimate of a single customer-service interaction moves from about $0.04 in 2023 to about $1.20 in 2026 — a thirty-fold rise in cost per completed task while the cost per token fell by orders of magnitude.

The budget consequences are already recorded. McKinsey's July 2026 Enterprise AI FinOps survey found 93% of respondents had exceeded their AI budgets, with a fifth limiting AI use on operating cost grounds. Reporting from late July put inference at roughly 85% of enterprise AI budgets, and cited Uber exhausting an annual AI budget by April after deploying a coding assistant to 5,000 engineers the previous December. This is Jevons in a quarterly reporting cycle: cheaper units, far more units, larger total.

V.The premium tier is the moat, and the open models know it

Academic work published in March 2026 anticipated exactly this split. Mingdeng Du of Wuhan University finds economy-tier token prices halving every 1.10 years and mid-tier every 1.55 years, while the flagship tier resists exponential decay entirely, carrying an average reasoning premium of 31.5×. The paper attributes essentially all cost reduction to software and architecture rather than GPU hardware, and records concentration falling sharply — a Herfindahl-Hirschman Index down 54.2%, from 4,558 to 2,086, crossing out of "highly concentrated" territory. Its own stated limits are real: 62 milestone observations and proxy-based market shares.

The open-weight challenge is closing the commodity gap and not yet the frontier one. DeepSeek V4 left preview on 13 August 2026 at $0.66/$1.98 per million off-peak for the Pro tier, scoring 80.6% on SWE-bench Verified — competitive on the older benchmark, but roughly 55% on the harder SWE-bench Pro against 64.3% for Claude Opus 4.7, with a NIST CAISI evaluation placing it about eight months behind leading US models. That eight-month gap is, for now, what the reasoning premium is priced on.

VI.The supply side is not behaving like a deflating industry

If falling token prices signalled falling demand, the hardware would show it. It does not. Nvidia's Q2 FY2027 results, released 26 August 2026, put revenue at $96.2 billion, up 106% year on year, with data centre at $89.0 billion, up 117%, gross margin at 75.0%, and third-quarter guidance of $108.0 billion. Jensen Huang's framing was that compute is now revenue.

The bear case is that this is precisely the setup for a reckoning. Allianz Research has put the gap between AI capex growth and sales growth at 46%, against the 32% divergence that preceded the 2001 telecom bust, and macro strategist Andreas Steno Larsen argued in early July that sustained token-price weakness would end the memory, hardware and data-centre trades for this cycle. Bloomberg's own coverage of that 20% drawdown was careful to note the ambiguity: a falling expenditure index is equally consistent with customers migrating to cheaper models as with demand weakening, and one index move is not a rerating. Citadel Securities reached a similar conclusion on 10 June 2026, forecasting uneven diffusion: frontier capability concentrating among well-capitalised firms with hard problems, everyone else settling into the cheap tier.

That distributional point is the one with macro consequences. PwC's 2026 work found the top 20% of companies capturing 74% of AI's economic value, only 12% of CEOs reporting both cost and revenue benefits, and 56% reporting no significant financial benefit yet. Cheap tokens have not, so far, broadened the winners' circle.

VII.How it plays out: three scenarios

Horizons are two to five years. Probabilities are our analytical judgment, not measurements, and are stated to make the reasoning falsifiable rather than to convey precision.