Anthropic released Claude Opus 5.5 on Tuesday at $4 per million input tokens and $20 per million output — a 20 percent cut from the Opus 5 it retires, with cached input dropping to $0.20 from $0.50. A million-token context window, multimodal input, and, on Anthropic’s own published numbers, better scores than Opus 5 across every benchmark run on both.

The same day, OpenAI shipped GPT-6 Sol and GPT-6 Luna at roughly half the per-token cost of the GPT-5.6 models they replace, as InfoWorld reported Wednesday morning. Two labs, one Tuesday, both marking down their best product on the day it ships.

The unanimous read is abundance: frontier intelligence is getting cheap, fast. Maybe. But nobody discounts the thing they expect to be superseded by their own next thing. You sell a flagship at full freight and let the successor arrive at a premium. A launch-day price cut is what you do when you’ve concluded this model will be the workhorse for a while — and when you know the other guy is doing the same math.

The Number That Matters Is the Invoice, Not the Benchmark

The benchmark claims are unverifiable in the usual way: they are the vendor’s own, published for both models being compared. BenchLM’s aggregator puts Opus 5.5 at 81.03 out of 100, fourth of 234 models, strongest in agentic tasks — while flagging that it lacks enough independent sourcing for a verified position. Fine. The interesting number is the one that lands on somebody’s invoice.

Anthropic is now selling its flagship at twice the price of Sonnet 5 ($2 in, $10 out) and four times Haiku 4.5 ($1/$5). A 2x premium for the best model the company makes is a thin moat. True, the same comparison table shows one model still asking $10 in and $50 out per million tokens — the top of the market hasn’t collapsed. That’s a rearguard, not a rebuttal.

There’s a second tell buried in the price sheet: a “fast mode” for Opus 5.5 at double the standard rate, returning output up to 2.5 times faster.

When You Charge for Latency, You’re Selling Time

Charging a premium for speed is what utilities, freight forwarders and cloud providers do. It’s the rate card of a business that has accepted the product itself is fungible and that the remaining axis of competition is when you get it. Labs are supposed to sell intelligence. Utilities sell time. Opus 5.5 is being sold both ways at once, and the second way is the one with the margins.

You can watch the same scramble up and down the stack. Google cut Gemini 3.7 Flash prices in mid-August. Nvidia was reported in late August to be circling Hugging Face with a $12.9 billion offer — a hardware company buying distribution. When the model layer commoditizes, everyone sprints to the ends of the chain.

An enterprise procurement lead at a mid-size insurer put it to me in a Slack message Wednesday morning: “Two of our three vendors called us before we called them. In six years of buying this stuff, that has never happened.”

A $965 Billion Valuation on Falling Unit Prices

Anthropic’s private valuation went from $183 billion at its September 2025 Series F to $380 billion in February 2026 to $965 billion on May 28, 2026. Run-rate revenue reached roughly $24 billion annualized by early 2026, up from about $3.7 billion in 2024. Those are not fake numbers, and the demand behind them is not fake.

But a company whose unit price falls with every release needs volume growth that outruns the price decline, permanently, just to stand still on the top line. That isn’t a software business. It’s a memory-chip business with better public relations.

Which means several constituencies have priors to revise. The people who argue compute scarcity guarantees pricing power — that proposition has now been tested twice in five weeks and failed twice. The people who want to treat these firms as strategic assets entitled to a durable rent: rents do not survive a price war. And the people who insist the whole thing is a bubble, who are half right for the wrong reason — a commodity collapse and a bubble pop look similar on a chart and point to opposite policies.

The Policy Built on a Moat That Just Got Shaved

Most of the current AI policy architecture — the export-control scaffolding, the sovereign-AI partnership push, the assumption that an American lead is a thing you can bank — rests on the premise that these labs have durable economics and that the frontier is a scarce asset controlled by two or three firms. Tuesday’s price sheets don’t refute that premise so much as strip its margin.

If the frontier is a cost curve, the relevant question stops being who has the smartest model and becomes who can build the cheapest unit. That’s a question about electricity, land, fabrication and industrial policy — not about guarding a secret.

Maybe the labs are wrong. Maybe Opus 6 lands in December and the price war reverses. But price is where a company tells you what it believes about its own product, and on Tuesday two of them said the same thing at the same time, unprompted.

Abundance is a fine thing to be handed. Just don’t mistake it for a moat.

Sources