OpenAI's Third Price Cut in a Month Beat Anthropic's Price. Check the Expiration Date First.

OpenAI cut the price of GPT-5.6 Sol, its flagship model, on August 21. Input dropped from $5 to $4 per million tokens, output from $30 to $20, a 20 to 33 percent reduction depending on which side of the request you're measuring. That is the third price cut to the GPT-5.6 family in four weeks, following July's cuts to Luna and Terra, and Enterprise DNA's tracking of the move frames it plainly as a response to Anthropic and cheaper Chinese open-weight models, not routine optimization. The new numbers put Sol at $4 and $20 against Claude Opus 5's $5 and $25, reported by Reuters and summarized by Digital Applied, the first time OpenAI's most expensive model has undercut Anthropic's on both input and output at once.

I wrote in early August that falling AI prices are the best news a business got this summer, and I stand by that. This latest cut is more of the same forcing function: competition, not a mandate, doing the work. But there is a detail in this specific cut that the "prices keep falling" headline skips past, and it matters if you are the one deciding what to build on.

The discount is not open-ended. Digital Applied's writeup of OpenAI's own model documentation notes OpenAI labels this "promotional pricing" and commits to it only "at least through November 21, 2026," three months out. That phrasing sets a floor on how long the price is guaranteed to last. It says nothing about what happens after.

Here is the steelman for not worrying about it. A business could reasonably say: who cares if it is a coupon. Three price cuts in four weeks means this is not a one-time event, it is a pattern, and the pattern itself is the durable thing. If OpenAI lets Sol's price creep back up on November 22, either a fourth cut follows within weeks because Anthropic or a Chinese lab forces the issue again, or a competitor is one email away from winning that workload. Under this reading, worrying about an individual promo's expiration date is missing the forest for the tree. The war is the constant. Grab the savings and move on.

I think that argument is mostly right and still misses something a business actually has to plan around: not whether prices keep falling in aggregate, which I agree they will, but what happens to the specific thing you built while a specific price was in effect. A frontier lab's promotional rate is revocable by contract, and the contract already tells you the date it can be revoked. An open-weight model you have downloaded and are running yourself is not revocable at all. DeepSeek, GLM, and Qwen's coding and reasoning models cannot expire on a date OpenAI's product team picks, because nobody controls the weights once you have them. That is the difference between a price war setting a temporary discount and a price war setting a floor, and I made the floor argument on this blog three weeks ago for a reason: it is the part of this trend a business can actually build a decision around.

So the practical read. If your workload is exploratory, a prototype, an internal tool you can swap the model under without much cost, ride the Sol discount while it lasts. It is real money and there is no reason to leave it on the table. But if you are making an architecture call that is expensive to unwind, a product feature tuned to one model's specific behavior, a vendor contract, a fine-tuning investment, do the math on what November 22 costs before you commit, not after. A 33 percent output discount that reverts is a 50 percent output cost increase from wherever you land, and that math should be in the spreadsheet on day one, not discovered on day ninety.

This is exactly the kind of decision we walk clients through: which parts of a workload should chase the cheapest current API price, and which parts are worth the extra week of engineering to run on something you own outright so a pricing page in Q4 can't move your costs for you. If you are building on a rate that has an expiration date and you have not run that math yet, tell me what you're running and I'll help you figure out which side of that line you're on.