Three model price moves in one week. Your AI SOW assumed none.

In the week of September 22, 2026, the price of running a frontier model moved three times, in three directions, from three vendors. Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20, and said it costs 40% less to run than Opus 5 on typical workloads ([Anthropic's announcement](https://www.anthropic.com/claude-opus-5-5)). OpenAI released GPT-6 Sol at $2 and $10 and GPT-6 Luna at $0.10 and $0.50, describing the change as reducing API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing ([OpenAI's announcement](https://openai.com/index/introducing-gpt-6-sol-and-luna/)). And Google's price list for Gemini 3.8 Flash now reads $0.75 per million input tokens through December 31, 2026, and $1.50 starting January 1, 2027, with output going from $3.75 to $7.50 on the same date ([Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing)).

Two cuts and one scheduled doubling, inside seven days. If you run an enterprise AI program, one of those three vendors is probably in your stack, and the price you are paying today is written into a contract that was signed when a different price was true.

Earlier this year I wrote about putting a model name in an SOW and watching that model get deprecated in March. Deprecation was the first thing AI broke in the way I had written services contracts for a decade. Price movement is the second, and it is the one most teams still have not written a clause for, because for most of the software era the unit price of a thing you bought did not change three times in a week.

Where the model price actually sits in a services contract

In an AI delivery SOW the model's price shows up in one of three places, and each one breaks differently when the price moves.

The first is a consumption pass-through. The client pays the vendor's token bill, usually through the integrator's account or a cloud marketplace, and the SOW says something like "model usage billed at cost." This is the honest structure and the one that handles a price cut best, because the saving flows straight to the client. It handles a price increase worst, because nobody budgeted for January and the first anyone hears of it is a finance question in February. The Gemini 3.8 Flash line is a scheduled, published, dated increase. A pass-through clause with no notice mechanism will still surprise the client, because the notice is sitting on a pricing page nobody on the program reads.

The second is a fixed fee with the model cost baked in. The integrator estimated tokens per transaction, multiplied by a price, added margin, and quoted a number. When the price falls, the integrator quietly keeps the difference and the client keeps paying the old economics. When the price rises, the integrator eats it, and the conversation about scope gets tense in a way that has nothing to do with scope. Neither side wrote down what happens, so both sides argue about fairness instead of reading a clause.

The third is outcome-based pricing, where the client pays per resolved case, per processed document, per closed ticket. This is where I have spent most of my time, and it is the structure that looks immune to model prices and is not. The unit economics behind the outcome price were built on a token cost. When that cost falls 40%, the integrator's margin widens and the client, who agreed to the outcome price precisely so they would not have to think about tokens, has no visibility into it. When the vendor doubles a rate, the integrator's unit economics can go negative on a price the client considers fixed.

The clause is not "pass-through." It is a notice, a re-test, and a split.

The mistake I see most often is treating price movement as a billing question. It is a change-control question that happens to carry a dollar sign. Three things need to be written down, and none of them are complicated.

The evidence gate a price cut has to clear

The strongest argument for moving to the cheaper model this week is the price. That is also the weakest reason to do it without a gate. In the operating model I use, AI earns decisions task by task, and a new model, however closely related to the old one, starts at the rung it can prove. Observe, suggest, act with approval, act with audit, act alone. A model that was acting alone on a narrow task last week does not inherit that permission because its successor is cheaper. It inherits it when the successor passes the same tests on the same job with the same audit trail.

That sounds slow, and it is slower than a config change. It is also the only version of a swap that a client can sign off on, that an auditor can reconstruct, and that a finance team can reconcile against the SOW. The three SOW clauses that model deprecation made mandatory apply here with one word changed: replace "retired" with "repriced" and the same machinery does the work ([the three SOW clauses model deprecation just made mandatory](/blog/ai-delivery-sow-model-deprecation-clauses)). A program that measures outputs but not outcomes will not even see whether the cheaper model changed anything that matters, which is the gap most pilots already have ([the eval gap](/blog/eval-gap-enterprise-ai-outputs-outcomes)).

The January line item

Of the three moves, the one that will cause the most trouble is the one that has not happened yet. Google's page states the Gemini 3.8 Flash rates through December 31, 2026, and the rates starting January 1, 2027, in the same cell. Anyone who reads the page knows. The programs that will be surprised are the ones where the price was read once, in a spreadsheet, in a quarter that already ended. Scheduled increases are the easiest contract event to handle and the most commonly missed, because the notice arrives before anyone is listening for it.

If your SOW has a model in it and no clause about what happens when its price moves, you do not have a pricing problem yet. You have a clause missing, and this week gave you three reasons to add it before the next one.

Frequently asked questions

What changed in model pricing in the week of September 22, 2026?

Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens and said it costs 40% less to run than Opus 5 on typical workloads. OpenAI released GPT-6 Sol at $2 and $10 and GPT-6 Luna at $0.10 and $0.50, a 50% reduction against their GPT-5.6 promotional pricing. Google's pricing page lists Gemini 3.8 Flash at $0.75 and $3.75 through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027.

Should we switch to the cheaper model right away?

Not on price alone. A change of model is a change under the SOW and should pass the same job-level acceptance tests as the model it replaces, with the same audit trail, before it takes over the same decisions. If it passes, the swap earns the saving. If it does not, the price was never the point.

What is the minimum clause to add to an AI SOW about model pricing?

Three parts: a notice obligation naming who monitors vendor pricing pages and how quickly they must report a change, a re-test requirement tying any model change to the existing acceptance tests, and a split that states in advance who keeps a price decrease and who carries an increase above a threshold. For consumption pass-through, add a monthly ceiling and a rule for introductory rates that expire on a date.