Alpha isn't in the model anymore. It's in the runtime. Two frontier labs — GPT-6.1 Sol and Gemini 4 Argon — published identical $2/$10 per-million-token pricing inside the same 48-hour window. That's not competition. That's a price signal. When two competitors converge on the exact same number, to the cent, the capability gap between them has collapsed into commodity territory, and the fight has already moved somewhere else: distribution, and the balance sheets that fund it. I've seen this setup before. In May 2022 I shorted LUNA into the Terra collapse — not because I understood algorithmic stablecoins better than the market, I didn't, but because the leverage was visible in the code and the order book before it was visible in the price. The agent war is the same trade, one layer up.
The battlefield is the agent runtime — the orchestration layer that lets a model actually do things: read your inbox, book flights, move money, call thousands of applications. Three deployment shapes are competing. OpenAI Dots runs every agent on its own dedicated cloud computer, covering more than 4,000 app integrations — a sandboxed VM plus tool-calling orchestration. Apple went the opposite direction: on-device processing stitched to Private Cloud Compute, trading raw capability for verifiable privacy. Meta's Muse chased scale with a free tier — 100 million tokens a week — riding it to 2.5 million US downloads. Anthropic isn't even in the consumer fight; it put $100M into Frontier Academy to train 10,000 engineers inside McKinsey, Accenture, and Deloitte by the end of 2027. That isn't a chatbot strategy. That's enterprise deployment packaged as consulting.
None of these labs shipped architecture, parameter counts, or context windows — the technical narrative is a distribution narrative. And the headline numbers are self-reported and unaudited.
The plumbing nobody wants to discuss is the money. OpenAI raised $11.1B in high-yield — junk-rated — bonds, the largest in company history, with a $30B SoftBank backstop. Anthropic carries a $51.8B compute commitment, 80% non-cancellable, against a $42B loss. A commitment approaching tens of times annual revenue, mostly irrevocable, funded to win a distribution bet that has to pay off before the capital runs dry.
Here's where the crypto-native reader should sit up. This is a leverage cycle. It just isn't denominated in a token you can short yet.
Run the free-tier math, because Meta's subsidy is the most mispriced number in the story. At the disclosed $2/M input and $10/M output, assume a favorable 3:1 blend — call it $4/M all-in. A single user burning the full 100M-token weekly allowance represents roughly $400 in API-equivalent cost per week, about $1,700 per month, against a premium tier priced at $20. The free tier only survives if real utilization sits under 5–10%. Meta isn't giving away AI. It's buying distribution with compute, and it can only do that because its ad cash flow — not its AI business — absorbs the subsidy. OpenAI runs the same trick with bond proceeds and SoftBank money. Apple buries the cost in hardware margin. The real variable in this war isn't model quality. It's whose parent cash flow can subsidize the longest.
Now the coupling the business press keeps missing: the agentic loop amplifies tokens. A single "daily task" can trigger dozens to hundreds of model calls. Where a chatbot consumes 10³ tokens per interaction, an agent completing a multi-step workflow consumes 10⁴ to 10⁶. That non-linear cost curve is what makes the free tier's ceiling so low and the $51.8B commitment look rational — someone has to lock capacity upstream before the demand curve even prints.
Strip the marketing and you get three unit-economics models that don't compete on the same axis. OpenAI's $500 Pro tier is a price anchor more than a revenue line — it makes the $20 and $100 tiers feel cheap while screening for high-value users, but a dedicated cloud computer per agent carries brutal marginal cost, so whether $500 even covers cost is an open question. Apple's "no subscription" isn't free either; the cost is bundled into hardware — an implicit subscription with the lowest acquisition cost and the most certain monetization of the five. And the standard war hiding underneath: those 4,000+ integrations are really an MCP-style tool-calling ecosystem. Whoever defines the function-calling standard owns OS-level power over the agent layer.
This is where on-chain rails enter. In 2025 I ran autonomous trading agents on Flashbots with $200,000 allocated — MEV-resistant execution, 10,000+ trades, 98% success, $45,000 net. The lesson wasn't that agents work. It's that agents are order flow. The moment an assistant can transact on a user's behalf, recommendation ranking becomes monetization, and the agent becomes a new intermediary between users and every application. That's a MEV-shaped opportunity and a MEV-shaped risk. The code doesn't care whether the caller is a human or a bot. The order book doesn't either.
The identical $2/$10 print deserves a second look. Two labs converging on the same price to the cent inside 24 hours is either commodity pricing or a coordination signal — the kind of thing FTC and DMA enforcers have historically flagged. And note the contradiction: the story frames $2/$10 as a commodity floor while reporting that both Google and OpenAI plan to double entry pricing. If the floor doubles, the commoditization thesis refutes itself. More likely it's capacity-constrained price discipline — or a live elasticity test. Restaking is leverage, but sleep is priceless; here the leverage is bonds and compute.
The consensus trade is simple: distribution wins, and Apple locks it down. I don't buy it. Apple's moat depends on the assistant staying default — and the DMA has already forced open default-app selection. The regulation the story treats as a geographic footnote is actually the thing dismantling the lock. And users multi-home. Nobody runs Siri or ChatGPT; they run both, routing system tasks to one and hard reasoning to the other. Low switching costs mean the agent market fragments into an oligopoly of three or four, not a monopoly. History agrees — smartphone OS, search, e-commerce all converged to oligopoly not single-winner markets.
Then there's the hole nobody is pricing: prompt injection. An agent that reads your email and holds tool-execution rights can be hijacked by malicious instructions embedded in that email. This isn't a chatbot saying something wrong. It's an agent moving money wrong. Every one of these products widens the attack surface, and none of the marketing mentions it. In a bull market, anyone can be a genius. The people who survive the next drawdown read the threat model.
Watch subsidy endurance, not benchmarks. The winner is whoever's parent cash flow outlasts the others' burn — and whoever converts personal context into a moat that's expensive to carry elsewhere. The 12-month signal: does the free tier shrink, does the $500 anchor hold, does anyone publish an audited utilization number? Trust the math, fear the hype, ignore the noise. The runtime war is a liquidity event. Position accordingly.