Contracting for AI While Prices Keep Falling

AI usage prices keep falling. Contracts signed in this environment need a different design from the ones signed a few years ago.

Part 1: Read It, Then Consider What It Means for Your Company

Through 2026, the major providers have cut prices one after another. OpenAI has reduced pricing on its higher-end models, and Anthropic is offering models with near-top-tier performance at roughly half the price band.

At the same time, the effort of switching has become smaller. Changing models used to mean rebuilding the calling code; increasingly it is a configuration change.

The significance of these two things happening together sits on the contract side.

A common arrangement among Japanese companies in the Philippines is to outsource development and contract for it as a package, including the models the contractor uses. There is nothing wrong with that arrangement in itself — the problem appears when a model number is written into the contract.

Six months later that model number is a previous generation, the price has fallen, and the contract does not permit a change. This situation does occur.

There is one more thing to read here. Prices falling means competition is fierce on the provider side too. A design that depends on a single provider is a risk on both price and supply.

Related: The Model Vendors Just Built Implementation Arms: What Local Subsidiaries Should Decide Now explains this in detail.

Part 2: Key Terms Explained (for Executives)

Token pricing is the unit on which AI usage is billed. You are charged according to volume processed. Whether a price revision reaches your invoice depends on how the contract is written.

Model specification is the centre of this. Writing "uses GPT-5.6" binds you to that model number. Writing "uses a model of equivalent or higher capability" accommodates generational change.

The migration clause is the procedure for changing provider or model. If the contractor can change it at their own discretion, it changes without your knowledge. If your approval is mandatory for everything, responses are slow. Rather than choosing one, split it by the type of change — that is the practical form.

Related: Lessons from OpenAI's Three-Tier GPT-5.6 Models: Designing AI Procurement That Matches Cost to the Job | Case Study for Japanese Companies in the Philippines explains this in detail.

Part 3: Applying This to Your Own Company

Settle three things before signing.

First, specify the model by capability rather than by number.

Writing "equivalent or higher capability than X" removes the need to redo the contract at every generational change. But without a separately agreed basis for judging "equivalent", interpretations will diverge later.

Second, write down how price revisions are handled.

Decide whether a provider's price cut reaches your invoice. It is not inherently wrong for the contractor to keep the difference, but if you do not state which it is, there will be a dispute later.

Third, confirm what stays with you after a switch.

I once outsourced a system implementation entirely, left the necessary conditions vague, and ended up with something that ran but could not be used. When it went well, I decided the initial design and the criteria for judgement myself, and left the implementation and the details of daily operation to the contractor.

If your own operating procedures and data design remain, you can migrate even after changing provider. If they exist only in the contractor's head, switching is not realistically available to you.

For reference, the large system I built for an online English school ran to a budget of over 10 million yen and three to five months of development. At that scale, whether you decided at the outset what connects to what determines directly how easily it can be modified afterwards. At the contract stage, confirm what is delivered as work product.

Related: Choosing in an Era When Switching Takes One Line: How Do You Judge an AI That Is "Strong in Japanese"? | A Case Study for Japanese Companies in the Philippines explains this in detail.

Part 4: Common Failure Patterns (What Not to Do)

Wrong 1: Writing a model number into the contract

Six months later it is a previous generation, the price has fallen, and you cannot change it.

Better: We wrote "uses a model of equivalent or higher capability" and set the basis for that judgement separately.

Wrong 2: Signing without deciding how price revisions are handled

Whether a provider's price cut reaches your invoice stays ambiguous, and becomes a negotiation later.

Better: We included one sentence at signing on how a revision would be treated.

Wrong 3: Leaving procedure documents out of the work product

The system runs, but why it was designed that way remains only with the contractor, so neither switching nor modification is possible.

Better: We named the operating procedures and judgement criteria as deliverables.

Wrong 4: Designing around a single provider

You end up with no options on either price or supply stability. Providers becoming unavailable does happen.

Better: We kept the calling layer replaceable and documented the switching procedure.

Three Tips for Getting Value

Tip 1: Keep contract terms short and adjust at renewal

Fixing a long term while prices are moving works against you. Renewing over shorter periods and matching the prevailing level is more practical.

Tip 2: Decide the basis for "equivalent or higher" first

If you specify by capability, decide what you will look at to judge equivalence. Without that, every renewal becomes an argument.

Tip 3: Rehearse a switch once

Actually moving to a different model surfaces assumptions that are not written in the contract. Do it once before you need to do it under pressure.

Bonus: How to Make Use of PH AI Works

PH AI Works supports Japanese companies considering entry into the Philippines, and Japanese business professionals who already have a local base, in putting AI and technology to work. On this topic of contracting and procurement, we can help in particular with the following.

• How to word a shift from model numbers to capability levels, and organising the basis for that judgement
• Checking what remains with your company after a switch, and identifying which documents belong in the deliverables
• Designing a replaceable calling layer and documenting the switching procedure

We offer a 30-minute free consultation for Japanese companies in the Philippines. It is fine to consult us before you have decided to adopt anything.

Sources

• OpenAI (official site; publishes its model line-up and pricing structure)
• Anthropic (developer of Claude; publishes its model line-up and pricing structure)
• Anthropic Newsroom (official news listing; model additions and pricing changes are announced here)

Note that pricing and the models on offer change frequently. This article reflects what has been published as of August 2026. Check each company's official pricing pages before making contractual decisions.