The discount on GPT-5.6 Sol expires on 21 November

A customer-service agent that costs 572 dollars a month in model spend today will cost 787 in January, without anyone touching a line of code. Not because of an announced increase, but because the price it was sized on was a promotion with its expiry date already written into the price list. That is the part of this week's price cuts that did not make the headlines. For anyone who has to get a budget approved, it is the only part that matters.

On 21 August OpenAI lowered the price of GPT-5.6 Sol for api use, the metered channel through which a model gets embedded into software you own. Reuters reported the move from 5 to 4 dollars per million input tokens and from 30 to 20 on output, lasting three months. It also noted that Pro, Plus and Business subscriptions keep their previous prices. The official documentation is more careful than the news coverage.

GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026.

That "at least" is everything a company has to go on regarding duration. On the other side Google is explicit. The Gemini api pricing page lists Gemini 3.7 Flash at 0.75 dollars input and 3.75 output through 31 December 2026, then 1.50 and 7.50 from 1 January 2027. The doubling is already a line item with a date next to it.

The bill for an agent handling eight thousand conversations

Let me build the most common scenario among those that come through trade associations, an agent answering questions about products, orders and deadlines. Eight thousand conversations a month, six exchanges per conversation, four thousand input tokens on each exchange across system instructions, retrieved documentation and conversation history, three hundred tokens of response. That comes to 192 million input tokens and 14.4 million output tokens a month. All of these are my assumptions. They are the fragile part of the calculation, because a context window of eight thousand tokens instead of four thousand doubles the largest line item.

At today's prices, short context and no caching, Sol costs 768 dollars of input plus 288 of output, so 1,056 dollars a month, 13 cents per conversation. If the price list returns to 5 and 30 on 22 November, the same activity costs 1,392 dollars, 17 cents per conversation, 32 per cent more.

Gemini 3.7 Flash on the same scenario without caching costs 144 dollars of input plus 54 of output, 198 dollars a month, 2.5 cents per conversation. From 1 January, 396, five cents.

Caching changes the proportions but not the direction. On Sol, tokens already seen in a previous request cost 0.40 dollars instead of 4, and in a customer-service agent the reused portion is nearly everything, instructions and documents. With 70 per cent of the context served from cache the bill drops to 572 dollars a month, 7.2 cents per conversation. After 21 November, assuming the cached rate stays at a tenth of the full price, the same configuration goes to 787 dollars. That is the increase in the first paragraph. It comes from the calendar alone.

What those numbers do not include

Three line items that disappear when you compare the two price lists and that weigh heavily in a real bill.

Long context. On Sol, beyond the standard window, the price becomes 8 dollars input and 30 output. Which means an agent carrying very long conversations or whole documents pays more than it paid before the cut. The promotion applies to the short window, and that is the distinction the headline does not make.

Data residency. The same OpenAI page shows a 10 per cent uplift on regional endpoints for models released on or after 5 March 2026. For a trade association handling its members' data that is not an option, it is the precondition for signing. The 572 becomes 629, and the 787 becomes 866.

Fast mode. Also from the price list, it doubles the rates. If someone switches it on to bring down the response time of a voice agent, the calculation has to be redone from scratch.

What I did not verify is Gemini's cached rate, or how Google prices context beyond its threshold, so the comparison above holds with caching off on both sides. And the 787 dollars in January hold if the price returns exactly to where it was, which is the most likely assumption but remains an assumption, because OpenAI wrote a start date and not an end date.

The flagship model is not worth five times as much

Here I take a position. For an agent that answers catalogue questions, books appointments, looks up an order and hands over to a person when the request falls outside its scope, the difference between 13 and 2.5 cents per conversation is not justified by the difference in quality. Those tasks are won on the quality of document retrieval and the precision of the instructions, not on the model's reasoning capacity. The flagship model earns its price where the task is open ended, code generation, analysis of contractual documents, long chains of steps in which an error halfway through propagates.

The serious objection is not about price. Switching models is not switching a line of configuration. Instructions tuned for one model behave differently on another, the format of tool calls changes, the way the model gives up when it does not know changes. Anyone who has done this twice knows the migration costs a few days of work and a set of test cases rebuilt from scratch. At eight thousand conversations a month those days pay for themselves quickly. At eight hundred conversations a month they do not pay for themselves at all. The reasonable choice is to stay where you are.

The only deadline that matters

Two numbers are needed before 21 November. The cost per conversation over the last thirty days, read off the invoice and not estimated. The same figure recalculated at the full list price. If the second exceeds what the company is willing to pay per conversation, the comparison test against a mid-tier model should be run in October, while there is still a month of margin, not in December with the invoice already grown.