On Tuesday OpenAI cut the price of its API roughly in half. We watched the announcement expecting it to translate straight into cheaper AI tools for our clients. It won’t, and the reason why is the more useful story.

What actually happened

OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, and priced them well below the generation they replace. Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. Luna, the smaller model most everyday tools run on, dropped to $0.10 and $0.50 per million tokens from $0.20 and $1.20. OpenAI also widened its discount on cached input tokens to 90%, which matters more than the headline cut for anything that reuses the same context repeatedly, like a support chatbot answering questions against the same product manual (OpenAI’s own announcement).

This is not a one-off. It’s the fourth or fifth time in two years that the raw cost of running a frontier AI model has roughly halved, and coverage of the release framed it as OpenAI defending its API market share against cheaper open-weight competitors (VentureBeat’s write-up has the full breakdown by model tier). If you build software on these models directly, as we do, this is unambiguously good news. If you pay a monthly subscription for a tool that happens to run on one of them, it is a lot less simple.

The number that doesn’t add up

Here’s the uncomfortable bit. Ramp tracks real AI spending across the businesses on its platform, and its AI Index shows token usage among businesses with connected AI grew 1,001% between January 2025 and April 2026. Over that same window, with per-token prices falling the whole time, total AI spend still grew 497% (Ramp’s AI token cost benchmarks). Prices falling by half sounds like it should shrink a bill. In practice it has barely dented one, because usage is growing far faster than prices are dropping.

Part of the reason is that businesses aren’t just using AI more, they’re using more expensive AI. Ramp’s data shows the share of spend going to premium, top-tier models rose from 5.7% in June 2025 to 55.9% in April 2026. Cheaper Luna-class models exist precisely so that routine tasks, like classifying an email or answering an FAQ, don’t need to run on the expensive model. Most businesses haven’t reorganised their AI use to take advantage of that yet, so a lot of routine work is quietly running on the priciest option by default, not by decision.

Why your subscription hasn’t moved

This is the part that actually affects most small businesses, because most of you aren’t calling OpenAI’s API directly. You’re paying a flat, per-seat price for a tool that calls it on your behalf, and that price is set by the vendor, not by the token market. Microsoft 365 Copilot in the UK currently runs from around £7.30 a user a month bundled into Business Standard up to £16.10 at list price for the standalone add-on (current UK pricing here), and ChatGPT Business costs $20 to $25 a seat a month plus VAT. Neither of those figures has moved because the model underneath them got cheaper this week, and there’s no reason to expect they will on their own. The API price is the vendor’s cost of doing business, not a promise passed on to you.

We’ve seen this pattern with every price cut since GPT-4 first came down in cost: the saving lands with whoever is closest to the model, and it stays there unless something forces it further down the chain. For a small business, that “something” is usually a competitor undercutting on price, not generosity.

What to actually do with this

None of this means ignore the price cut, it means look in the right place for the saving. Three moves are worth ten minutes each this week.

First, separate what you’re paying for by the seat from what you’re paying for by usage. A per-seat subscription (Copilot, ChatGPT Business, most AI-enabled SaaS add-ons) won’t get cheaper on its own. A pay-as-you-go API bill, or a tool that passes API costs through to you, should be getting cheaper right now, and it’s worth checking your last invoice against this week’s prices to see if it has.

Second, if you or a developer you work with are building anything in-house on the API directly, an FAQ bot, an inbox triager, a document summariser, check whether it’s still pointed at last generation’s model. Repointing it at Luna instead of Sol, where the task allows it, is often a five-minute change and a 90%+ cost reduction on that workload, not a 50% one.

Third, if a vendor tries to sell you an “AI is included” price rise on the back of this news, ask them directly what their underlying model cost just did. The honest answer, this week, is that it fell. A subscription price that only ever goes up, on a product whose core input cost just halved for the fourth time, is worth a harder look before you renew.

We build AI tools directly on top of these models for clients precisely so the savings above pass through rather than disappearing into someone else’s margin. If you want a second opinion on whether what you’re paying for AI right now still makes sense, that’s a conversation worth having before your next renewal, not after.