The AI Price in Your Budget Is Really Two Prices

“AI is getting cheaper” is the kind of line that slides through a budget review without anyone stopping it. It’s half right. The AI price in your budget is really two prices, and they’re moving in opposite directions.

One is the price of using an AI model, and Anthropic just cut it hard for its small one. The other is the price of the memory chips that AI servers, laptops and phones all draw from, and by TrendForce’s account it’s still rising. They don’t land on the same line of your budget, so each needs its own estimate. But they correct the same assumption, so both belong in the room before the 2027 plan is signed.

The price of using a small model fell

On October 7, 2026, Anthropic released Claude Haiku 5.5, the small model it designed for “high-volume, cost-sensitive tasks.” Models are priced by the token, a small chunk of text, and Haiku 5.5 costs $0.10 per million tokens going in and $0.50 per million coming out, for prompts up to 100,000 tokens. Its predecessor, Haiku 4.5, charged $1 and $5. That’s a tenth of the price. Above 100,000 tokens the rate steps up to $0.50 and $2.50, five times the new short-prompt rate and still half the old one.

Anthropic’s own summary is that the model “now costs around 75% less to run” on average. The footnote matters, though. That average comes from Anthropic’s mix of requests, where 90% of requests to Haiku 4.5 fell under the 100,000-token line, and it accounts for a new tokenizer, the way the model splits text into tokens, that in Anthropic’s words “uses slightly more tokens per task.”

So the cut is real, and it’s on a public price list. What it saves you depends on how long your prompts are, how much the model writes back, and whether a small model does your work well enough to keep.

The memory that AI servers compete for

Every laptop, phone and server runs on memory chips, and memory makers decide how to split their factories between them. On September 30, 2026, the research firm TrendForce published its forecast under a blunt headline: AI server demand sustains memory contract price increases. Its explanation for phone memory is that “strong demand and higher margins in AI-related applications are prompting suppliers to continue shifting capacity toward server products, leaving mobile DRAM allocations tight.” It projects contract prices for conventional DRAM, the working memory inside devices, to rise another 10 to 15% in the last quarter of 2026 over the quarter before, and it says the memory supply for PCs “could decline in 2027 as more capacity shifts to server applications.”

One phone makes it concrete. In August 2026, TrendForce estimated that the 256 GB iPhone 18 Pro’s parts would cost about 38% more than its 2025 predecessor’s, “driven by soaring memory prices,” and that memory went from about 10% of the parts cost of the 256 GB iPhone Pro a year earlier to about 34% in the third quarter of 2026. It also said Apple may offset some of that by accepting lower margins.

A parts cost isn’t your invoice, and a research firm’s estimate isn’t a price you’ve been quoted. But some of the pressure on the hardware you buy runs through the same factories that feed AI, the physical layer we looked at in the chips, power and cables underneath AI. It’s the part of AI’s cost that never appears on a model’s price page, and a company that buys no AI at all still sits downstream of it.

Before anyone signs the 2027 budget

For the 2027 budget, we’d run three checks. Each ends in a decision you can act on, and the first comes first for one reason: it shows how much of your budget each price can touch, and whether the second check is worth running.

1. Find out which lines each price can reach. Keep the budget categories you already have, and look at three of them: AI your developers call directly and pay for per token, AI you pay for as a fixed subscription, and equipment that runs on memory. The Haiku cut can reach only the first line, and only for work you move to Haiku 5.5. A per-seat subscription is priced on its own terms, so assume no saving there unless your vendor confirms a lower charge, and it helps to know what each AI contract actually counts. If your per-token spend is small, the price cut can’t save you much next year, and you can go straight to the third check.

2. Test before you book the saving. Start with the usage records, because they show which tasks carry enough spend to be worth testing. Among narrow, repeatable tasks with a clear quality bar, take the one carrying the most per-token spend, and have the person who runs your AI tools pull how many tokens go in and out and how many prompts run past 100,000 tokens. Then they run the cheaper model on a sample of those real tasks against the quality bar you use today, and compare the cost of each finished task, counting retries and the time people spend fixing the output. Count the work of switching once, as a one-time cost. Keep a lower bill apart from staff time freed up, and project both at the workload you expect next year. If the cheaper model holds up, pilot it. If it doesn’t, keep what you have. Until a test shows a saving, there’s no saving in the budget. And a cheaper model can raise this line instead of lowering it. Anthropic says Haiku 5.5 suits narrowly scoped tasks “that might otherwise have been cost-prohibitive,” so work that was too expensive may become worth doing. If you take any on, budget it as new spend with its own case, not as a saving.

3. Cost the devices you need, on their own terms. This check runs whatever the second one finds, because a cheaper model in the cloud doesn’t change whether a laptop is still getting security updates. Build the replacement list from security support and day-to-day need, then price that list at today’s quotes instead of carrying over last year’s figures. In August 2026, TrendForce said memory prices had risen “five- to sevenfold since the beginning of 2025,” which is reason enough to refresh last year’s estimates. A higher quote alone shouldn’t decide whether you replace or wait. The output is one recommendation from whoever owns the device budget: what to replace in 2027, what that costs at today’s quotes, and when each quote should be refreshed.

What the two numbers leave open

TrendForce’s figures are a research firm’s estimates and forecasts, not company disclosures, and a component cost isn’t a retail price, especially when a maker like Apple may absorb part of it. Haiku 5.5 is one vendor’s launch, and other models and providers price differently. Neither TrendForce release cited here says when memory prices will ease, and TrendForce expects the increases to slow from earlier quarters in the last quarter of 2026, not to stop. And neither number tells any single company what its budget will look like, which is exactly why the checks start from your own budget lines and usage records.

What the checks leave you with is three decisions instead of one hopeful number: which budget lines each price can touch, whether to pilot the cheaper model or keep the current one, and what the devices you need cost at today’s quotes. AI now reaches a company’s budget through two doors, the price of the model and the memory it competes for, and only one of them is announced on a launch page.