Four Labs Repriced in Seven Days. The On-Chain Version of That Trade Does Not Clear.

The Document Had Decimals and No Links

In late September 2026, a text began circulating that described Gemini 4 entering post-training. It carried index scores to one decimal place. It carried version numbers down to the patch level — v4.3.2. It carried dates. It carried the name of a new senior vice president. It carried no primary link to any of it.

That was the first thing I noticed. It was not the interesting thing.

The interesting thing was the window around it. Inside seven days, four frontier labs repriced their models. That is measurable. Repricing leaves a mark on a system I can audit. A leaked engineering milestone does not.

Within forty-eight hours of the document moving through group chats, a basket of tokens claiming exposure to AI compute, decentralized inference, and agent frameworks traded volume that had no relationship to any on-chain cash flow I could locate. I pulled the transfer graphs. I clustered the sender and receiver sets. The buying was concentrated in a small number of addresses, most of which had received their inventory inside the preceding three weeks, and most of which had never interacted with the protocols they were bidding up.

Volume is noise. Intent is signal. And the intent behind that buying had nothing to do with whether Gemini 4 finishes post-training.

The ledger lies; the code tells. So I went to the code — and to the arithmetic, which is the only part of this story that cannot be narrated into a different shape.

What the Document Claims, and Why the Claims Cannot Be Taken at Face Value

The text makes a specific set of assertions. Gemini 4 has entered post-training. Gemini 3.5 Pro has been stuck in partner testing since June. Gemini 3.6 Flash scores 34.0 on an Intelligence Index v4.3.2; Opus 5.5 scores 57.6; GPT-6 Astra scores 52.7; GPT-6 Sol lands somewhere in the forties. Grok 4.7 is priced at $2 input / $6 output per million tokens and is given no score at all. Opus 5.5 is anchored at $4 / $20. GPT-6 Sol and Luna carry a "permanent" 50% discount. A researcher named Kavukcuoglu has been elevated to SVP above Gemini development. The venue for disclosure was an offline summit hosted by The Information.

I want to be precise about my own position before I dissect any of that. My verifiable knowledge does not extend to 2026 frontier models, to an Intelligence Index v4.3.2, or to any of the version numbers above. That is not a hedge. It is a boundary, and boundaries matter because most of what gets written about AI right now is written by people who do not know where theirs are.

What I can do is read the artifact as an artifact. And the artifact has defects that are more informative than its contents.

The provenance chain is broken in the text itself. In one place the source is labeled as unclear. In another it is described as a blockchain and Web3 news outlet. A Web3 outlet publishing pure frontier-model analysis is not impossible, but it is a category mismatch — the kind that content-aggregation pipelines produce routinely and that primary reporting almost never does. The numbers are precise to a degree that implies measurement and are accompanied by nothing that would allow measurement to be checked. The internal framing wobbles: one passage describes a four-lab race, the closing passage describes a three-way frontier race. And the model naming — Astra, Sol, Luna, Gemini 3.5 Pro, 3.6 Flash, 4 — follows the exact cadence that generative text defaults to when it is simulating a product roadmap.

I have read a lot of whitepapers that were written to be believed rather than to be true. In 2017, as a high school junior, I reverse-engineered the TON preliminary distribution schedule in Python and found that sixty percent of the supply sat with insiders while the document described decentralization. The tell was not any single sentence. The tell was that the arithmetic did not close when you ran it forward.

This document has the same tell, at the level of provenance rather than supply. Treat it as scenario narration — market text — not as reporting. Its analytical value is in what it reveals about expectations, not in what it asserts about facts.

Which raises the question the crypto market did not ask. If the underlying claim cannot be verified, why did anything trade?

The Measurement Problem: You Cannot Settle a Contract on a Number You Cannot Audit

The composite index score is the load-bearing element of the entire narrative, and it is the weakest.

A single number that compresses coding, reasoning, tool use, long-context retrieval, and multilingual performance into one decimal is not a measurement. It is a marketing artifact with a decimal point. It answers the question "which model should I feel best about" and not the question "which model should handle this specific workload at this specific cost." For a developer deciding on a routing table, the composite is actively harmful, because the composite hides exactly the variance that matters.

Then there is the comparison problem. The text places Gemini 3.6 Flash at 34.0 against Opus 5.5 at 57.6 and GPT-6 Astra at 52.7. But Flash, in Google's product line, has always designated the low-cost, high-throughput tier. It is not the flagship. The flagship, per the text's own account, is Gemini 3.5 Pro, which has been sitting in partner testing since June and has no score in the document at all.

So the document is comparing a light tier against competitor flagships, then reporting a gap of 13.5 to 24.0 points, then letting the reader convert that into "Google is 40% behind." The forty percent figure only holds against the single highest scorer. That is not analysis. That is reference-frame selection, and it produces an artifact, not a finding.

Here is the part that matters for anyone building on-chain. A metric that cannot be independently reproduced cannot be used as an oracle input. This is the oldest problem in the space and it has not been solved by better dashboards; it has been buried under them.

Consider what people are trying to build with this class of data. Model-release prediction markets. Tokenized compute indices. Routing contracts that rebalance automatically when a new model ships. Every one of those instruments requires a number that a neutral party can recompute from a public specification. None of them can use an index that exists only inside a narrative.

I ran into the identical wall in 2021, when I clustered wallet addresses on OpenSea and found fifteen interconnected wallets wash-trading Bored Ape inventory to inflate the floor by roughly two million dollars. The floor price was published. It was precise. It was also fabricated, because the trades that produced it had no economic substance. The volume chart was not a measurement of demand. It was a measurement of how cheap it was to move a number.

Composite model indices and manipulated floors fail in the same way. Both are numbers that lots of capital depends on and that almost nobody can verify. The distance between a published number and a reproducible number is where all the risk lives.

The Compute Arithmetic: Two Months Cannot Do What the Text Says It Did

This is where I stop reading and start calculating, because arithmetic does not negotiate.

The text implies that Gemini 4 was announced as starting pre-training around July 21 and had entered post-training by late September. Two months. Say the training run was in the 10^26 FLOPs class — the threshold that triggers dual-use reporting obligations under the US executive order framework that was in force in 2023 and has been politically unstable since. Set up the cluster the way frontier clusters are actually set up: on the order of 100,000 accelerators, dense BF16 throughput near 1.9e14 FLOPs per second per chip, model FLOPs utilization at a realistic 40%.

Effective cluster throughput lands near 7.6e18 FLOPs per second. Two months is about 5.2e6 seconds. Multiply.

You get roughly 3.9e25 FLOPs.

That is a tenth of the 10^26 threshold, from a run the document insists was significant in scale. There are three ways to reconcile this, and only three.

One: the actual compute requirement is in the 10^25 range, which maps to a dense-equivalent parameter count somewhere around 500 billion to a trillion. That is large. It is not obviously "significantly larger" in a way that is differentiating in 2026.

Two: the cluster is far bigger than 100,000 accelerators. To hit 10^26 in two months you need something like 250,000 chips running at the same efficiency. That is possible for exactly one or two entities on the planet, and it would be the single biggest capex disclosure of the year. It appears nowhere in the document.

Three: the two-month window is not a pre-training window. It is the gap between an announcement and a status update, and the text has read "we are starting" as "we started." Engineering organizations announce program kickoffs and report program milestones months later. Conflating the two is the most common error in AI timeline coverage, and it is the one I would bet on here.

I gave this calculation a D-confidence grade when I ran it, because it rests on assumed utilization and assumed cluster size. But the direction of the conclusion is robust even under generous assumptions. A two-month pre-training cycle does not produce a frontier-scale model unless the cluster is twice the size anyone has publicly described.

The document never engages with any of this. No architecture. No parameter count. No data composition. No synthesis ratio. No context length. No pricing. It describes a program without describing a single input to it. Silence is the first red flag, and this document is mostly silence with numbers on top.

Where the Decentralized Compute Trade Actually Breaks

Now the part the token market skipped.

Decentralized compute networks — the Render-class, Akash-class, io.net-class of protocols — have spent several years positioning themselves as the alternative supply layer for AI training and inference. The pricing war narrative is presented as a tailwind for them. Cheaper models, more usage, more demand for compute, more demand for the tokens. That is the pitch.

Friction reveals the true structure. Run the structure.

Frontier pre-training does not rent spot GPUs from a permissionless marketplace. It requires co-located, high-bandwidth-interconnect clusters with consistent hardware, deterministic failure behavior, and multi-month reservation windows. No decentralized network on earth offers that today at frontier scale. The interconnection topology alone disqualifies it. Decentralized supply is competitive for embarrassingly parallel, latency-tolerant workloads — rendering, batch inference, fine-tuning of small models, data preprocessing.

So the training side of the AI boom does not touch these networks. What touches them is inference. And inference demand is the variable that the agent narrative actually moves.

Here is the tension the document creates without noticing it. Long-horizon agents consume tokens in a fundamentally different pattern than chat. Multi-step planning, tool calls, reflection, retries. A single user-visible task can burn one to two orders of magnitude more tokens than a single-turn completion. If that is the direction, then even with unit prices falling, total inference volume can rise — a classic Jevons argument, and probably correct as far as it goes.

But decentralized inference supply is not priced off total volume. It is priced off the marginal cost of the cheapest credible alternative. And the cheapest credible alternative is the hyperscaler with self-designed silicon, depreciating its own capex on its own schedule. That is where the arithmetic turns hostile. A self-designed accelerator amortized internally produces a different cost floor than a merchant GPU bought at market price with a vendor margin on top.

The decentralized compute thesis does not fail because the technology does not work. It fails because the price floor is set by an entity that owns the silicon. Follow the gas, not the hype — and in this case, follow the depreciation schedule.

The Pricing Collapse Is Real, and the Language Around It Is Worse Than Wrong

The one part of the document I believe directionally is the price compression, because price compression is the one thing in frontier AI that has been consistent for three years.

Output pricing anchoring: Opus 5.5 at $4 / $20 against a historical Claude 3 Opus anchor of $15 / $75. Grok 4.7 at $2 / $6 against a Grok 3 anchor around $3 / $15. GPT-6 Sol and Luna carrying a "permanent" 50% discount off an undisclosed base. Across the frontier tier, output cost per million tokens has moved from the $75 range toward $6–20. That is a compression of roughly four to twelve times in about eighteen months. I do not need to verify the specific SKUs to accept the direction. The specific SKUs are decoration. The curve is the finding.

Now the language. The document describes a "coordinated pricing structure maintained under the Amodei pacing framework" that "collapsed in less than two days."

That phrasing is a legal category error, and it is not a small one.

Coordinated pricing between competitors is per se unlawful under US antitrust law and under EU competition law. No set of frontier labs is maintaining a public coordination structure and describing it as such in writing. What happens in an oligopoly with four or five participants and high transparency is something else with a similar outward shape: parallel pricing. Participants observe each other's published rack rates, respond within days, and converge on similar structures without any agreement. That equilibrium is legal, unstable, and breaks whenever one participant decides the volume is worth more than the margin.

The difference between collusion and parallel pricing is the difference between an agreement and an observation. It is the entire difference. If the document's framing were quoted in a regulatory filing or a private action, it would do real damage — not because the underlying conduct was illegal, but because someone described legal conduct using the vocabulary of illegal conduct.

I flag this because crypto has the same disease at higher amplitude. People call market-maker inventory management "manipulation." They call aggressive-but-legal listing behavior "insider trading." They call a large treasury position "a rug in progress." Sometimes they are right. Usually they are using the wrong category, which means when the real thing happens — and it does happen, regularly — the accusation carries no weight because the vocabulary has been spent on things that were not crimes.

My 2021 wash-trading work held up publicly because I never used the word manipulation. I showed fifteen wallets, the transfer graph, the timing pattern, and the resulting floor distortion. Algorithmic truth requires no defense. Fifteen clustered addresses executing offsetting trades on the same inventory is fraud by construction, and I did not need to characterize it. The data characterized itself.

The document's "coordinated pricing" claim does not characterize itself. It asserts. That is the difference between evidence and argument, and it is why I grade the commercial section of this document no higher than C despite agreeing with its direction.

The Agent Turn Is the Real Story, and It Is an On-Chain Security Story

Buried under the pricing material is the only structurally new claim in the document. Gemini 4's stated priority is coding and long-horizon autonomous agents. That is the sentence worth spending time on, because agent capability changes the risk surface in a way that has nothing to do with benchmark scores.

Here is the mechanism. A chatbot's failure mode is a wrong output. It is bad, it is visible, and a human reads it before anything happens. An agent's failure mode is an executed action. It is also bad, but it is not visible in the same way, and in many configurations there is no human in the loop to catch it.

Prompt injection is the clean example. In a chat context, an injection attack leaks a system prompt — annoying, occasionally embarrassing, recoverable. In an agent context, an injection attack causes an unauthorized tool call. Now attach a wallet.

Now the failure mode is a signed transaction.

I do not need to speculate about the severity here, because I have already run this class of simulation. In 2022, after the Terra collapse, I rebuilt the UST mechanism in a local sandbox and replayed the death spiral under thin-liquidity conditions. The finding was not that the team behaved badly. The finding was that the peg mechanism was structurally unable to hold when liquidity left, and no amount of community sentiment changes a bonding curve. I wrote five hundred words and assigned no blame, and it landed harder with developers than anything moralizing would have.

The methodology transfers directly. Do not audit an agent for the happy path. Audit it under adversarial input.

Run the adversary. An agent holds session credentials, API keys, and in the on-chain configuration, signing authority. An injected instruction that reaches the tool-call layer can move funds. There is no chargeback. There is no fraud department. Settlement is final by design, which is the entire point of the rail and also the entire problem when the counterparty is a language model under adversarial influence.

Hallucination compounds this rather than sitting beside it. In a single-turn interaction, a hallucinated fact is corrected by the next message. In a hundred-step agent workflow, a hallucinated intermediate state propagates. Step forty assumes step thirty-nine's output was valid. By step ninety, the agent has executed a sequence of actions that are individually plausible and collectively wrong, and some of them touched money.

The engineering answers are not governance answers. Minimum-permission credential scoping. Session keys with hard expiry. Spending caps enforced at the contract layer rather than the prompt layer. Explicit human confirmation at the transaction boundary — not for every step, but for every step that moves value or grants access. Deterministic allowlists on tool invocation.

Every one of those is a system design property. None of them is a policy statement. You cannot prompt your way out of an authorization problem. The labs will publish safety frameworks. Frameworks do not enforce. Contracts enforce.

The Middle Layer Is Where the Value Actually Accrues

If four labs sit within a workable band of capability while pricing and per-domain strength diverge, then the rational developer does not pick a model. The rational developer routes.

That creates a demand for three things, and only one of them is well served today. Dynamic routing that selects a model per task type. Cost observability — the measurement of unit-task cost, which is not the same as unit-token price. And evaluation decomposed from a composite score into per-domain plus per-cost rankings.

The unit-task metric is the important one and it changes the economics more than the price war does. Unit-task cost equals unit price divided by token efficiency for that task class. A model that costs twice as much per token but consumes a third as many tokens per completed task is the cheaper model. That single relationship restores pricing power to efficient models and punctures the pure price-war logic. It also means raw token price — the number everyone quotes — is close to meaningless for predicting total spend.

There is an on-chain version of this and it is mostly unbuilt. Verifiable inference, routing markets, cost-attestation feeds. All of them require the same primitive: a cost-per-task figure that a third party can reproduce. That is an oracle problem wearing new clothes, and the oracle problem has never been solved by adding a token. It has been solved by reducing the number of things that need to be trusted.

I have watched this pattern enough times to be bored by it. The thing that gets tokenized is rarely the thing that captures the value. Routing and cost measurement are genuinely valuable layers. Whether a token can sit on top of either one and retain that value is a separate question, and the answer is usually no, because the layer's customers are developers who will simply call an API.

What the Bulls Actually Got Right

The bull case deserves a fair stress test, because parts of it hold and dismissing it wholesale is its own kind of sloppiness.

First, the price collapse. Directionally real, and I said so above. The specific SKUs in the document are unverifiable, but four-to-twelve-times output compression over eighteen months is consistent with everything observable in API pricing history. If you have been modeling AI cost curves, you already had something like this in the spreadsheet.

Second, cost-per-task as the unit of account. This is the correct reframing and it is genuinely bullish for small, well-optimized, open-weight models — and by extension for on-chain inference where the operator controls the runtime and can tune the token budget directly. A composite index ranking does not capture this. Cost-per-task does. Give the bulls that point, because it is the strongest one they have.

Third, the direction of capability. Coding and long-horizon agents are the highest-value, most defensible demand in the market. Coding has an objective quality signal — tests pass or they do not. Agents produce measurable task completion. Neither of those is a vibe, which separates them from most of what gets called AI progress.

Fourth, and this is the one the crypto side actually cares about: an agent economy implies machine-to-machine payments. Not conceptually — mechanically. An agent that books, buys, provisions, and settles needs a payment rail that does not require a human to approve a wire. That is a real requirement, and programmable money is the most plausible answer to it. This is the only segment of the AI-crypto bridge that is not narrative.

Now the counter. The bullish case for compute and inference tokens requires token consumption to grow faster than unit prices fall. Jevons. It might hold. But it requires agents to actually close tasks rather than demo them, at scale, in production, with money attached. And nobody — not the document, not the token market, not a single AI-adjacent project I have reviewed this quarter — has published a credible unit-task cost curve. Every bullish model I have seen is a straight line drawn through a hope.

The Failure Mode Nobody Prices

One structural point about Google that the document omits entirely, and its omission is the most revealing thing in the whole text.

Google is the only entity in the comparison that owns its accelerator design, its cloud, its distribution channels, and a cash-generating core business. That combination has one specific consequence in a price war: it can sustain a lower price longer than anyone funded by external capital. At a $2 / $6 anchor, a firm paying merchant margin on compute and a firm amortizing its own silicon are not in the same business.

But the implication for capital is the opposite of what people assume. Google's failure mode is not death. It is chronic marginalization — staying in the market, shipping models, retaining a developer base, and never quite winning the routing table. There is no清算 event. There is no blowup. There is a slow decay in share and relevance that never produces a single headline.

That is exactly the shape of a governance token that never dies and never pays. It trades. It has holders. It has a treasury. It has a roadmap. It has no cash flow to the holder and no terminal event to force a reckoning. The absence of a清算 moment is not stability. It is the mechanism by which capital stays trapped.

Incentives align, or they break. And here, the incentive to hold is decoupled from any mechanism that returns value. That is true for a governance token, and it would be true for a persistently-third-place model business if it were publicly traded as a pure bet.

What to Watch, and What to Ignore

Ignore the composite score. It will be published, it will be quoted, it will move tokens, and it will be unverifiable in exactly the way this document's scores are unverifiable. A number that cannot be reproduced is a number that can be produced.

Ignore "permanent" pricing language. Pricing is a commercial decision with no legal force. The word permanent in a release is a signal about competitive anxiety, not about cost structure. When a firm announces a permanent cut, the honest reading is that its differentiation is no longer sufficient to support the old price.

Watch the absence. Gemini 3.5 Pro has reportedly been stuck in partner testing since June. That is a longer delay than any technical problem in a frontier program typically produces. Partner testing is a process gate, not a research gate. Three months of process friction in an organization that has infinite compute says more about internal decision-making than any benchmark number will.

Also watch who is silent. A document that names a newly elevated SVP and does not mention the CEO of the lab that SVP works under is telling you something about organizational change, or telling you the document was generated. Either way, the silence is data.

The number to actually watch is not an index score. It is the first audited, reproducible cost-per-task benchmark published by a party with no model to sell. When that exists, routing becomes allocatable, capital can price it, and the AI-adjacent token complex gets its first real anchor in something other than narrative. Until then, everything trading on this story is trading on a decimal point with no link behind it.

So here is the question. When the next document like this circulates — and it will, next quarter, with better formatting and more precise decimals — what will you actually be holding when the window closes?

History is just data waiting to be read. Most people read the headline instead.