AI Got 33% Cheaper Last Week. Most Companies Will Draw the Wrong Conclusion.

Last week, OpenAI cut the price of its top coding model by close to a third. Three weeks before that, it cut the price of another model in the same family by eighty percent. If part of your AI business case rests on this month's token rate, it is already out of date.

What actually happened

On August 21, OpenAI dropped API pricing for GPT‑5.6 Sol: input tokens fell twenty percent, from five dollars to four dollars per million, and output tokens fell a third, from thirty dollars to twenty. The promotion runs through November 21, and OpenAI has been clear that it is a temporary window, not a new list price. It follows a July 30 repricing of two other models in the same family: Terra dropped twenty percent, Luna dropped eighty percent. Three price moves on one model family in six weeks.

This is not an OpenAI only story. Google cut prices when it launched Gemini 3.7 Flash. Anthropic has been undercutting rates to win enterprise accounts. Every major lab is fighting the same war on the same front: whoever is cheapest per token wins the volume, and volume funds the next round of training. For the companies selling AI, price is a weapon. For the companies buying it, that should change how you think about cost.

The strategic reframe

Most companies read a headline like this and land on one of two conclusions: wait, prices will keep falling, or lock in now before they change again. Both miss the point.

The real issue is that token pricing was never a stable enough number to build a strategy on. It moves every few weeks, in either direction, for reasons that have nothing to do with your business. If your adoption decision hinges on this month's rate card, you are not running a strategy. You are trying to time a market you do not control, using pricing you cannot predict, set by companies whose priorities are not yours.

What holds up regardless of what a provider charges next month is the workflow underneath it: what task you are automating, what quality bar you need, who owns the output, and what it is actually worth to your business when it works. Price is one input to that calculation. It should never be the calculation itself.

What this actually means for an SME

Build your business case on outcome value, not on input cost. If a workflow saves your team four hours a week, that value does not change when a vendor's price does. Price affects your margin on the win, not whether the win is real.

Design your workflows and prompts to stay as independent of the provider as you reasonably can. If moving from one model to a cheaper one in the same family, or switching providers entirely, means rebuilding your prompts and retesting your outputs from scratch, you have built a dependency, not a workflow.

Treat a price cut as a trigger to revisit scope, not as a trigger to adopt. A task that was too costly to automate at five dollars per million tokens might be worth automating at four. That is a legitimate reason to look again. It is not a reason to have waited in the first place.

Count the real cost, not the sticker price. A cheaper model that needs more prompting, more correction, and more of your team's oversight time is not actually cheaper. The token price is usually the smallest line in an AI budget. The people time around it is usually the largest.

None of this means ignore pricing. It means stop treating it as strategy. The companies still using AI well two years from now will not be the ones who caught the best rate this quarter. They will be the ones whose workflows were solid enough that the rate barely mattered.

Try this week

Pull one AI tool's usage report from the last thirty days. Calculate cost per completed task, not cost per subscription or per token. If you cannot produce that number in under ten minutes, that is the gap worth closing before the next price change, not after.

Sources

• OpenAI Cuts GPT-5.6 Sol API Prices by Up to 33% Through November 21 — WinBuzzer, August 23, 2026
• Advancing the price-performance frontier with GPT-5.6 — OpenAI, July 30, 2026