The Cheap Model Now Costs More Than the Expensive One

OpenAI’s GPT-6.1 Sol has cost $2 and $10 per million tokens since 30 September. On Thursday the same model, unretrained, also went on sale at $12 and $60 — six times the price for lower latency, and, by OpenAI’s own note, about 1.2 times what the flagship it undercuts costs to run normally. No weights moved. A dial was added, and the cheap model became the dear one. The unit on sale this week was not a model but a route to one, and in most cases the buyer is not the one choosing it.

The buying question has changed shape. For two years it was which model is best; this week it became which one answered, at what effort, over which wire — and who decided. Four dials now sit between a request and a bill: tier, effort level, latency, and whether it ran on your machine or in a data centre. Two move at the vendor’s discretion, per request, unreported.

The model’s name has stopped predicting what it costs, and this week it stopped by a factor of about 286. Seven days produced list prices from $0.042 per million input tokens — Microsoft-Decision-1, live on Friday, which writes no text and only scores options it is handed, output free — up to $12 for Sol run fast. Not the same job, which is the point: the work is now unbundled into bands so far apart that “what does a model cost” has no answer, only a configuration. Google’s new effort levels make capability a slider, and its own documentation admits the slider is metered.

The unit on sale this week was not a model but a route to one, and in most cases the buyer is not the one choosing it.

The dispatcher was the week’s real product launch, and it belongs to the seller. Microsoft’s hybrid intelligence for Windows, set out on Wednesday, routes work between models on the PC and in the cloud to cut token spend, with the split meant to be invisible to the user. Decision-1 sells the same insight as a part: the routing, classifying and should-I-continue judging an agent does all day, which until now paid flagship rates to a model built for prose. Xbox Research, Microsoft says, ran 10,000-odd pieces of player feedback through it at a two-hundredth of GPT-6 Sol’s price. The saving is real. The receipt is not.

The cheap tier got good enough that withholding it became a commercial decision, not a technical one. Claude Haiku 5.5 arrived on Wednesday having gone from 0.0% to 39.2% on Terminal-Bench 4.0 at a tenth of its predecessor’s short-prompt list price, and Anthropic published the arithmetic against itself: the real saving is nearer 75% than 90%. Two days later Google’s free tier dropped to Flash-Lite alone. The floor is rising fast, and the gate is coming down onto it.

Yes, but: this is also what a market growing up looks like. Electricity is priced by time of day, cloud compute by instance class, and nobody calls that deceit. Cohere, whose flagship agent release this week was mostly ways to stop agents spending, says why: a year of large deployments found enterprises sacrificing one of security, capability and cost control to keep the other two. Routing really is cheaper, and cheap is how a small operation affords this at all. But a kilowatt-hour is a kilowatt-hour. What varies here is the quality of the answer, not the quantity of a uniform good — and on Saturday Satya Nadella published a framework demanding human-readable evidence for every meaningful action a model takes, three days after his own company shipped a router designed not to say which model did the work.

Consequences and opportunities

• A benchmark score now describes a configuration, not a product. If effort and route vary per request, the published number came from a setting you cannot be sure you bought.
• “We use GPT-6” stops meaning much in a procurement document. A contract naming a model but no route, effort level or duty to report which ran specifies nothing enforceable.
• The free tier becomes a different product, not a smaller one. Since Friday, Gemini’s unpaid users get a weaker model, not less of the good one — far harder to notice than a smaller quota.
• Silent routing and auditability pull against each other. Every control on Nadella’s list assumes you know what ran; a quiet dispatcher is the cheapest way to break that.
• Opportunity: cheap competence is the week’s real gift to small operators. Haiku 5.5 at $0.10 per million input tokens clearing 39.2% of Terminal-Bench was flagship-priced work a month ago; this site’s budget rests on that arithmetic.
• Opportunity: the local half of the week suits Georgia better than the frontier half. Frontier access starts at $19.99 a month and is climbing; a good small model for a language with little text online is something an institute that just trained 27 researchers and got a server can own outright.

What to watch

• Whether any vendor starts showing which model answered. Windows’ router does not, and nothing announced this week suggests one is coming.
• Whether a lab publishes benchmark scores per effort level and route rather than per model name. The first to do it makes everyone else’s numbers look vague.
• Whether Ultrafast’s six-times premium survives its first month, or latency settles into a tier like every other dial.

Ask what your AI costs and the only honest answer now opens with a question: which one answered?