What AI model price cuts mean for your AI budget and vendor contracts

OpenAI and Anthropic cut the prices of several AI models between July 30 and September 29, 2026, including an 80% cut for OpenAI's GPT-5.6 Luna, as Max Woolf laid out on October 25. A healthcare organization can count on that saving only where its contracts pass lower model costs through, so check how each AI vendor sets its price before you build the 2027 budget.

Below are the cuts, why they may not reach your bill, and six price terms to check before you sign or renew an AI contract.

What did OpenAI and Anthropic cut?

Both companies price their models per million tokens, the units of text a model reads (input) and writes (output), with a lower price for cache reads, "where the model reuses context it has already processed," in Anthropic's words. Three of the 2026 changes show the size of the cuts, and OpenAI's developer changelog records the two OpenAI ones.

Starting July 30, OpenAI says, GPT-5.6 Luna "costs 80% less." Its model page now lists $0.20 per million input tokens and $1.20 per million output tokens, and Woolf gives the launch price as $1 and $6. On August 21, OpenAI cut GPT-5.6 Sol to $4 per million input tokens and $20 per million output tokens, which it called "20% lower input pricing and 33% lower output pricing," as a promotional price. Its newer GPT-6.1 Sol, released September 29, lists at $2 and $10. On September 22, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5, with cache reads at $0.20 per million tokens, 60% less than Opus 5.

Anthropic says its tests show that at default settings Opus 5.5 "will cost 40% less than Opus 5 on typical workloads," because it costs less per token and "uses fewer tokens per task." That is the model maker's own estimate for its own model.

Will the cuts lower what you pay?

Only if your contract makes the vendor pass them on, or the vendor chooses to. If an AI product you buy runs on a model from OpenAI or Anthropic, the vendor pays that model maker for what it uses, at whatever rate it has agreed, and you pay what your contract says, whether that is per user, per visit, per month or per use. For example, if a vendor charges a flat fee per clinician, its model costs can fall by half while the fee stays where it is, unless the vendor cuts it or the contract lets you revisit it. If the vendor passes usage fees through, your bill moves with the model maker's price, down or up.

Some cuts are temporary. OpenAI's pricing page says GPT-5.6 Sol's promotional price "is available at least through November 21, 2026." The same page says OpenAI charges 10% more for its regional processing endpoints, which keep data in a chosen region, on models released on or after March 5, 2026. If your vendor uses OpenAI's U.S. region to meet your data rules, ask whether that 10% is in its price.

Will your AI spending fall?

Not necessarily. Vercel, which runs a gateway that software teams use to reach many AI models, reported on October 15, 2026 that estimated spending through the gateway, at list prices, was about 13 times higher in September 2026 than in December 2025, while the average price per token fell 47%. Part of that fall came from customers moving work to cheaper models.

The median paying customer active in both months more than doubled its estimated spending. These are software teams using one company's service, so the figures say nothing direct about healthcare organizations. They do show that a lower price per token and higher total spending can arrive together, so budget for how much your staff will use a tool, as well as for what each use costs.

What should the contract say?

Settle these six terms before you sign or renew. They apply whether the vendor sells you a finished product or access to a model.

Plan terms change in both directions. When Anthropic released Opus 5.5, it also raised the five-hour usage limits on its Pro, Max, Team and seat-based Enterprise plans, so subscribers can do more within each five-hour window. Woolf reports that when OpenAI announced GPT-6.1 Sol, it cut the usage allowance on the $200-a-month tier of Codex, its coding tool, by 50%, which OpenAI said reflects its more efficient models. Ask your vendor to state in the contract what each plan includes, so a change in either direction shows up in writing.

For the rest of the cost, how to score an AI vendor before the demo works through a build-versus-buy example in which hosting and model fees together are one line of a build's yearly cost, next to the staff time to own it. The exit terms that protect your data, such as return or destruction when the contract ends, are in what happens to your data when an AI vendor shuts down.

Who can add up a vendor's full cost?

The AI Vendor Evaluation adds up each shortlisted vendor's full cost over three years, tests each one on cases from your own organization, and ends with the contract terms to ask for before you sign, on data use, exit rights and service levels. It takes two weeks and costs $7,500 for up to three vendors or $9,500 for four or five.