Two price movements this year point in opposite directions. The software you buy by the seat got more expensive. The intelligence inside it got dramatically cheaper. Most businesses have noticed the first and none of the second, which is how you end up paying more for something whose underlying cost has collapsed.
01 — What happened
On 4 December 2025 Microsoft told its partners that commercial Microsoft 365 suites would carry a global price update across all purchasing channels effective 1 July 2026. New and renewing customers pay the new prices from that date; existing customers see them at their first renewal afterwards. Microsoft’s own licensing FAQ attributes the increase to “the significant innovation delivered over the last several years and the expanded value customers will gain with new additions to the suites, including major advancements in security and IT management.”
That is one vendor, but not an outlier. Gartner’s July 2026 forecast, reported by TechTarget in August, put worldwide software spending growth at 15.5% for the year against total IT growth of 14.2%. The same piece quotes the plainer mechanism: “Software spend, on average, every time any contract comes up, seems to go up by 10%, 15%.”
Now the other direction. On 22 September 2026 Anthropic released Claude Opus 5.5 and cut its API price 20%, from $5 to $4 per million input tokens and $25 to $20 per million output; cache reads fell 60%, from 50 cents to 20. OpenAI answered the same day with two cheaper models: GPT-6 Sol at $2 input and $10 output per million tokens, and GPT-6 Luna at 10 cents input and 50 cents output. Both vendors publish these rates themselves. On OpenAI’s current pricing page, Luna’s input price is one hundredth of GPT-6 Astra’s $10.
So in one calendar year the price of a seat rose by low double digits, while the price of a unit of machine work fell by anywhere from 20% to an order of magnitude.
02 — Why it matters
Those two lines measure different things, and the gap between them is the useful information.
What you pay for a per-seat subscription is a bundle: the capability, plus the interface, the integrations, the compliance paperwork, the support contract and the vendor’s cost of selling to you. The capability is now the cheapest ingredient in that bundle, and getting cheaper every quarter. Everything else is holding or rising.
That has a practical consequence. List price used to be a rough signal of what something cost to deliver. It no longer is: the published model tiers alone span a hundredfold range, so two products at the same price per seat can differ by orders of magnitude in what the capability inside them costs to run. The difference goes to packaging and distribution — a legitimate thing to buy, as long as you know that is what you are buying.
Larger organisations have begun to act on it. McKinsey’s State of AI 2026 survey, reported in September, found 32% had decided against buying off-the-shelf software and built their own with agentic coding tools instead — rising to close to half among the firms attributing at least 5% of their EBIT to AI.
03 — The economics
Put the two sides in the same units.
Microsoft’s December partner announcement lists Copilot Business at $21 per user per month, on top of an underlying Business Basic, Standard or Premium licence. For an illustrative 25-person business that is $525 a month, or $6,300 a year, to add AI assistance to seats you already pay for.
Now price one specific workflow directly. Take the same business handling, say, 1,000 inbound enquiries a month. Assume each enquiry consumes roughly 4,000 input tokens — the message, the customer context, the instructions — and produces about 600 output tokens: a classification, a short summary and a drafted reply. That is 4 million input and 0.6 million output tokens a month.
- On GPT-6 Luna at $0.10 and $0.50: about 70 cents a month.
- On GPT-6 Sol or Claude Sonnet 5.5 at $2 and $10: about $14 a month.
- On Claude Opus 5.5 at $4 and $20: about $28 a month.
The honest caveat first, because it is the point: these are not the same purchase. Copilot sits inside the tools twenty-five people use all day, improving a thousand small unnamed tasks. The $14 does one named task well and nothing else.
But the ratio is roughly forty to one, and a ratio that size changes which question matters. It is no longer “which product is better value per seat” but “which work do I actually want done, and how often”. If you can name the work, you can now buy it at close to its marginal cost. If you cannot, you buy a bundle and hope the right part gets used.
Three costs sit on the build side, and they are why this is a narrow conclusion rather than a licence to build everything.
Specification. Someone has to decide what good output looks like, in writing, with edge cases. That is the real work and the real cost, and no price cut touches it.
Integration. The model is useless without the enquiry, the customer record and somewhere to put the answer. Connecting those three almost always costs more than the inference.
Ownership. A field changes upstream and the thing breaks quietly. And where the run cost is metered rather than fixed, a build gets more expensive as it succeeds — the shift we covered when vendors moved to charging per resolution rather than per seat.
Inference at 70 cents a month does not make a build cheap. It makes inference stop being the reason not to.
04 — What businesses should do
List every per-seat subscription with its seat count and renewal date. Then mark any where fewer than half the seats open it in a given week. Those are the lines where you pay for packaging you do not use — usually the easiest money in the building.
Know what your renewal costs before the invoice arrives. On commercial Microsoft 365 suites the new prices apply at your first renewal after 1 July 2026, not on a fixed date, so the number has probably not reached you yet. Find out now, while there is still time to change the shape of the subscription rather than just approve it.
Split the stack into two piles. Systems of record — accounts, payroll, the CRM — and workflow glue, meaning the subscriptions you bought to move information between the things in the first pile. Buy the first pile. The second is where the build case lives.
Interrogate every per-seat AI add-on with one question: which task, how often. Priced per seat, a feature used by three people out of twenty-five costs what it would if all twenty-five used it. If the honest answer is “we are not sure yet”, buy it for the three.
If you have a build candidate, pick the one with volume and stable rules. Several hundred repetitions a month, a process unchanged in a year, one named owner. Write down the number you expect to move before anyone writes code — because hours saved is the wrong way to value it.
05 — Where technology fits
For most businesses reading this, the honest answer is not a build.
Systems of record should be bought. The economics of building your own accounting or CRM platform have not improved; the compliance surface alone settles it. If your records live in spreadsheets, the first move is a connected system rather than anything custom — the pattern behind outgrowing spreadsheets.
The next most common answer is an integration, not an application. Where three subscriptions exist because two systems do not talk, the cheapest fix is to make them talk and cancel the third.
A genuine build case is narrow: one high-volume workflow, no off-the-shelf product that fits your operating model, and a per-seat licence covering a product only a handful of people touch. That shape is real and increasingly common. It is not most workflows, and the same discipline applies as to deciding against a mobile app — the cheap thing to build is rarely the valuable thing to build.
And the legitimate no: if your software bill is modest, nothing is duplicated and every subscription is used, there is nothing here to optimise. Falling token prices are not a reason to start a project. Do nothing, and revisit at renewal.
06 — RevenueStack’s take
Our view: the important development is not that AI got cheap. It is that the price of a capability and the price of the product containing it have come apart far enough that list price has stopped carrying information.
The vendors understand this perfectly well. Microsoft’s own partner guidance frames the July increase as “a key moment” to “accelerate renewals and upgrades” and says the change “opens the door for upsell motions” — written for the channel rather than for customers, and unusually candid about what a price rise is for. Nothing wrong with it. It is the other side of the table explaining how the table works.
The discipline this calls for is not “build instead of buy”. It is narrower and duller: be able to name the work. Businesses that can say which tasks they want done, how often, and to what standard will spend less over the next three years, because they can buy that work close to its marginal cost. Those that cannot will keep buying bundles priced on the assumption that nobody is counting.
Counting is the whole advantage. It usually takes an afternoon and a spreadsheet, which is the least glamorous part of the process and reliably the highest-return one.
Sources
- Pricing and packaging updates for commercial Microsoft 365 suites coming in 2026 — December 2025 Partner Center announcements — Microsoft, 2025-12-04
- Microsoft 365 Packaging and Pricing Updates Public FAQ — Microsoft, 2026-10-05
- Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models — SiliconANGLE, 2026-09-22
- Claude pricing — Anthropic, 2026-10-05
- API pricing — OpenAI, 2026-10-05
- Gartner predicts 14.2% increase in IT spending: How to prepare — TechTarget, 2026-08-12
- The Build-vs-Buy Shift: 32% of Enterprises Bet on Agentic Coding Tools — Forkast News, 2026-09-01
