The frontier stopped getting cheaper

I.Two launches, one price

On 1 September 2026 Anthropic shipped Claude Fable 5.1 and Mythos 5.1 at $10 per million input tokens and $50 per million output. Two or three days later — sources differ on whether it was the 3rd or the 4th — OpenAI shipped GPT-6 Astra at $10 and $50. Not close. The same.

That is a new fact, and a narrow one, so it is worth stating precisely what it is and is not. It is not evidence of collusion; identical prices emerge routinely in markets with two credible sellers, similar cost structures and public price lists. It is evidence that neither lab believes it can win frontier customers by undercutting the other, and that both now expect to compete on capability and on the structure of the bill rather than on the headline rate. OpenAI, as one pricing review put it, stopped competing on price at the top.

The gap to everything below is the part that should reorganise a budget. Google’s Gemini 3.8 Flash, out 2 September, lists at $0.75 / $3.75. Meta’s Muse Spark 1.3 came in at $1.25 / $4.25. DeepSeek’s V4.1-Flash, released 11 September with a 1M-token window and a 552-billion-parameter mixture-of-experts design, charges $0.15 / $0.60 off-peak. The market has not converged on a price. It has split into two markets.

II.Competition moved into the fine print

Price competition did not stop. It moved to levers that are conditional, revocable and much harder to put in a spreadsheet.

Anthropic held Fable’s base rate flat and cut cache reads from $1.00 to $0.25 — a 75% reduction that the company estimates makes typical workloads about 25% cheaper and heavily agentic ones up to roughly 45% cheaper. That is a real cut, and it only lands if your workload is cache-heavy. DeepSeek prices by clock: the $0.15 / $0.60 rate applies off-peak and doubles during peak windows. Google’s $0.75 / $3.75 is an introductory rate that expires on 31 December 2026 and then doubles to $1.50 / $7.50. OpenAI’s cheaper GPT-5.6 Sol tier is a promotional rate guaranteed only through 21 November 2026.

Astra adds surcharges in the other direction. Cross 272,000 input tokens and the entire request reprices at double the input rate and 1.5× output — a 280K-token request costs about $7.10 where the same work trimmed to 272K costs roughly $3.72, per CloudZero’s worked example. Low-latency “fast” mode doubles the standard rate again. The headline is a 1.05-million-token context window; the affordable part of it ends at about a quarter of that.

III.The index is still falling, and that is the trap

The strongest case against this brief is that prices are plainly still dropping, and it deserves a fair hearing. BenchLM’s frontier price index stood at 16 on 11 September 2026 against a March 2023 baseline of 100 — an 84% decline in three and a half years. Independent trackers put the fall further still when measured against capability rather than calendar.

But the aggregate hides a split that matters more than the average. TokenCost’s price index, which tracks every major published price change since March 2023, separates the two: the cheapest model capable of a given task fell by roughly 200–300×, while the best available model fell only about 12×, from $30 to $2.50 per million. The widely quoted “300× cheaper” figure is a statement about the commodity tier, not the frontier. And momentum has now gone flat: BenchLM recorded a 0% month-over-month move in that same frontier index, cautioning that there is no guarantee of a straight-line decline.

A second caveat cuts against the thesis honestly. Astra is priced 2.5× Sol but appears to consume fewer tokens per agentic task, so part of the premium is recovered in efficiency. Per-token price was never the right unit; cost per resolved outcome is. That is exactly why the flat frontier list price is a warning rather than a curiosity — the one number enterprises can forecast has stopped moving, and the numbers that replaced it are workload-dependent, promotional, or both.

IV.The bill lands on the workflow, not the token

Gartner’s August forecast is the clearest statement of the mechanism. Routing a task to an agentic reasoning model rather than a chatbot raises inference cost at least fivefold, and cost per agentic workflow is forecast to rise more than fivefold through 2028 even as token prices fall. Efficiency gains are being spent, not banked: cheaper tokens make more expensive models affordable, and more capable models are given longer-horizon work that consumes far more tokens.

Trace that through the obvious channels. On productivity, the arithmetic of automation no longer improves automatically with time — a task that is uneconomic at $50 per million output tokens today is not reliably economic in eighteen months, so the pilot that failed its business case does not get rescued by waiting. On market structure, a stable frontier price and a collapsing commodity price push work down-tier: the rational architecture is a cheap model doing most of the volume, escalating to the frontier only where it demonstrably pays, which is precisely the “inference tiering” Gartner recommends. On capital flows, the frontier premium is what services the buildout — Futurum put 2026 hyperscaler capex at $660–690 billion across five firms in its February analysis, nearly double 2025.

V.Why the floor is being defended now

The timing is not incidental. Anthropic is reported to have selected Nasdaq for a potential October listing at a valuation approaching $2 trillion, alongside a $50 billion US datacentre programme announced with Fluidstack in November 2025. SoftBank closed an upsized $11.9 billion loan on 14 September to fund its OpenAI position. A company weeks from a roadshow does not cut gross margin on its flagship product, and a lender syndicate does not underwrite a price war.

This is judgment rather than fact, and should be read as such: the identical $10/$50 point looks less like a competitive equilibrium than a jointly defended floor, held while both labs convert capability leads into revenue visibility for capital markets. The conditional discounts are how they keep large customers without touching the number an analyst will model.

VI.The escape valve is open, and it runs through Hangzhou

What stops a defended floor is a credible substitute, and one arrived four days after Astra. DeepSeek V4.1-Flash posts 74.2 on DeepSWE v1.1 — against 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol on the same benchmark — at $0.60 per million output tokens off-peak. If that result survives independent replication on tasks that matter, the frontier premium is being charged for reliability, tooling and enterprise assurance rather than raw capability, and it is vulnerable.

Two forces push the other way. Regulation raises the cost of being a serious vendor: on 1 September the European Commission sent its first AI Act information requests to more than 30 providers, covering safety and copyright — compliance overhead that favours incumbents. And compute is contracted years ahead at scale; reporting this week identified Anthropic as the buyer in a $13.7 billion six-year GPU agreement, a single-sourced claim that should be treated as unconfirmed. Fixed costs of that size are an argument for holding price, not cutting it.

VII.How it plays out: three scenarios

Probabilities below are analytical judgment, not measured likelihoods, on a horizon running to end-2028.