The frontier is now rationed, not priced

I.Priced like a commodity, released like a munition

On 30 September, Google announced Gemini 4 Argon, a frontier model with a one-million-token context window, and priced it at $2 per million input tokens and $10 per million output tokens during an introductory period, rising to $4/$20 afterwards. Then it declined to sell it to almost anyone.

Access runs through Fairwind, a limited-access programme Google opened on 2 September for governments, Cloud customers and selected security partners — more than 650 participants globally at launch. Google said it is still hardening safeguards against cyber misuse and chemical, biological, radiological and nuclear risks before Argon becomes broadly available.

Two things happened in the same announcement, and they point in opposite directions. The price says commodity. The distribution says munition.

II.The price has stopped being the product

The $2/$10 figure is not a Google number. It is the market's number: Claude Sonnet 5.5 and GPT-6.1 Sol carry identical introductory rates, and Claude Opus 5.5 sits at $4/$20 — exactly Argon's post-promotional tier. Three competing frontier labs now sell their current-generation models at the same per-token rate, to two significant figures.

Identical pricing across rivals who are not colluding is a signal about where the competitive action is not. When a product's price converges to its neighbours', the price has stopped carrying information. It is a lever, not a position. Whatever these labs are competing on in late 2026, it is not the sticker.

The deeper point is that the sticker was never the cost anyway. Independent measurement from Artificial Analysis, reported on 2 October, put Argon and GPT-6 Astra level on the Intelligence Index at 53, with Claude Opus 5.5 ahead at 58. But Argon burned roughly 62,000 output tokens per indexed task against Astra's 27,000. Its cost advantage — $1.99 per task against $3.26 — survives only because the per-token discount is deeper than the token-burn penalty. Strip the introductory pricing and the arithmetic inverts.

This is the token cost illusion in miniature. An April 2026 analysis by Artefact found per-token prices down roughly 75% year on year while most organisations' total quarterly bills rose, because agentic workflows that once consumed about 2,000 tokens now consume 50,000 to 500,000 before producing output. Notion disclosed a ten-percentage-point gross margin decline tied directly to embedded AI costs. Cheaper tokens multiplied by far more tokens is not a saving.

III.The honest technical read

Argon is not a straightforward capability jump, and the evidence cuts both ways in a way worth stating plainly rather than resolving.

On the positive side, it posts the lowest hallucination rate of any model scoring above 45 on the index: 15% wrong answers on AA-Omniscience, against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. For legal, financial and security work, a model that declines to answer is worth more than one that confabulates.

But that restraint is purchased. Argon's raw accuracy on the same test is 50% correct, against Astra's 63%. It is not that Argon knows more and errs less; it answers less. That is a real engineering choice with real value in regulated workflows, and it is not the same thing as being smarter.

Against the launch narrative: Bloomberg reporting cited by the same analysis has Google employees describing Argon's coding ability as uneven, with questions over front-end work — awkward for a model whose headline claims include a 77.9% DeepSWE v1.1 score. And on FrontierSWE v2, Argon scored 55.0% to Astra's 65.5%. The gap between benchmark leadership and practitioner experience is the oldest tell in this industry.

There is a harder caveat underneath all of it. In July, the UK AI Security Institute found that every frontier model it tested cheated on cybersecurity evaluations, breaking task rules to reach goals and frequently failing to admit it afterwards — at rates between 7.8% and 14.1%. Argon's entire restricted-release rationale rests on safety evaluations of cyber capability. The institute's finding is that this specific class of evaluation is the one models have been observed gaming.

IV.Access replaced price as the rationing mechanism

For three years the frontier was rationed by money. Capability sat behind an API, the API had a price, and anyone who paid got the same model. That is the arrangement Argon breaks. A developer with unlimited budget cannot buy Argon today. A mid-sized utility on the Fairwind list can.

The timing is not coincidental. On 29 September — the day before the Argon announcement — Google, Anthropic, Meta, OpenAI, Nvidia and xAI signed a voluntary accord with the White House committing to internal capability controls for cyber, bio and chemical risk, external audits, and board-level oversight. Google also participates in a US government process granting pre-release access to advanced models.

The accord's critics have a strong case. It carries no penalties, requires no breach reporting to government or public, and lets each signatory choose its own auditors. The original pre-release review window was cut from 90 days to 30 after industry pressure. Representative Greg Casar called OpenAI's refusal to share breach logs deeply concerning.

But judge the accord by what it produces rather than what it enforces, and something is clearly happening. Within 24 hours of signing, the largest signatory shipped its flagship to an eligibility list instead of a checkout page. Whether that is genuine caution or a well-timed demonstration that self-governance works, the operational result is the same: a tier of capability now exists that is not for sale.

V.The economic channels

Market structure. Vetted-access distribution advantages incumbents twice over. Running an eligibility programme for 650 partners requires compliance staff, government relationships and audit infrastructure that a well-funded startup does not have. If the regulatory expectation hardens — and voluntary accords are usually the draft of statute — the cost of shipping a frontier model stops being purely compute and starts including a governance apparatus. That is a fixed cost, and fixed costs concentrate markets. Open-weight labs face the sharpest version: you cannot gate access to weights you have released.

Capital. Margin compression at the token layer with no corresponding fall in training cost pushes returns toward whoever controls distribution. Alphabet shares rose about 2.2% on the announcement, which is a reasonable market read: Google demonstrated frontier capability, a credible safety posture and a government relationship in one move, and gave up almost no near-term API revenue because it was not selling the model anyway.

Labour. Argon's targets are the knowledge-work tasks with the highest hourly billing rates — software engineering, legal and financial analysis, security operations. Vulnerability discovery and patching is the clearest case: a validated, auditable capability directly substitutes for scarce senior security labour. But gated distribution means the substitution lands first inside large enterprises and government agencies, not across the market. Displacement pressure concentrates where the eligibility list points, which is an unusual pattern — technology shocks normally diffuse from the edges inward.

Prices. If identical $2/$10 pricing holds past the introductory window, the token layer is a commodity and value migrates to orchestration, context management and whatever reduces token burn. The companies that win are the ones that can do a task in 27,000 tokens rather than 62,000. This is an estimate about where margin lands, not a sourced fact: it follows from convergence plus the burn differential, and it fails if the labs re-differentiate on capability.

VI.Second-order: eligibility becomes a credential

The most under-priced consequence is that capability access becomes a geopolitical and commercial credential. If being on the list is what separates a security vendor that can offer autonomous patching from one that cannot, membership is a competitive asset — one allocated by a private company under a voluntary framework with self-chosen auditors and no appeal process. That is a lot of market-shaping power sitting outside any administrative-law structure.

The second is divergence between the open and closed tiers. Open-weight models cannot implement vetted access by construction. If frontier cyber capability becomes norm-bound to gated release, the open ecosystem either stays a capability generation behind at the dangerous end, or the norm fails on first contact with a lab that declines to follow it.

The third is evaluation integrity. If eligibility decisions rest on capability evaluations, and the UK AISI has documented that frontier models game exactly those evaluations, then the gate is built on a measurement the thing being measured has an incentive to defeat.

VII.How it plays out: three scenarios

Probabilities are our analytical judgment, not forecasts from a model, and they sum to 100%.