Skip to content
Spark
Writing

Adoption and strategy

AI just got 75% cheaper. Your business case barely moved.

Anthropic’s new small model costs around 75% less to run than the last one, and OpenAI cut its prices too. Run the numbers on a real workflow and the saving is $27 a month. The business case lives somewhere else.

By Seb Lawson · 8 October 2026 · 5 min read

Model prices fell hard this autumn, and for most professional firms it changes almost nothing. On a typical document workflow the model bill was already a rounding error. The money is in the hours of human review, and those did not get any cheaper.

The headlines were real. Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, against $1 and $5 for Haiku 4.5, and says it costs around 75% less to run. OpenAI put GPT-6 Sol and Luna into its API on 22 September at half the promotional prices of the models they replace, and began rolling GPT-6 out in ChatGPT on 7 October. For scale, a token is roughly three quarters of a word.

Run the numbers on a real workflow

Take a trust company processing 2,000 KYC documents a month: passports, utility bills, company registers, bank letters. Assume each one is about 10,000 tokens in and 1,000 tokens out once the fields are extracted. Those are our assumptions, not vendor figures, and your documents will vary.

On Haiku 5.5, that is 20 million input tokens at $0.10 per million and two million output tokens at $0.50 per million. Three dollars a month. On Haiku 4.5 it was $30. The price cut saved $27.

Now the people. Keying those documents by hand at six minutes each is 200 hours a month. With extraction and a 90-second human check on every document, it is 50 hours. That is 150 hours saved, and at a loaded cost of £45 an hour, £6,750 a month. The price cut is worth $27 a month. The workflow is worth £6,750.

That gap is the whole argument. Model fees were around 1% of this workflow’s running cost before the cut. Now they are about a tenth of one percent. No business case should have hinged on them, and none should change because of them.

What cheap models do change

Two things, both useful. First, coverage. When reading a document costs a fraction of a penny, you stop sampling. Every inbound email can be triaged, every statement read, every file checked against the checklist, not just the ones someone had time for. You go from a sample to 100%.

Second, a second opinion. At these prices you can afford a second model to check the first, with disagreements flagged for a person to resolve. On the workflow above that doubles a $3 bill. It is the cheapest control you will ever buy.

One caution. Switching to whichever model is cheapest this month means a new vendor, a new data flow and a new entry in your AI register. Price is a reason to test a model, never a reason to skip the due diligence. Our data protection guide covers what to check.

The cost line that does grow: seats

Per-seat licences are where AI spend quietly creeps. Microsoft 365 Copilot Business lists at $21 per user per month on an annual plan, with an $18 promotional rate until the end of December. A 30-person firm at list price pays $630 a month, $7,560 a year, before anyone builds an agent.

Agents are metered on top, in Copilot Credits at $0.01 each on pay-as-you-go. Microsoft’s billing rules include some employee-facing agent use in the Microsoft 365 Copilot licence when the agent runs under the user’s own identity. Anything triggered another way, on a schedule or by an incoming email, consumes credits. That is a variable cost most firms have not budgeted for, and it is the one to watch. Check how your own licence is treated before you build.

A seat is not expensive. At a loaded cost of £45 an hour, a $21 seat pays for itself if it saves its holder less than an hour a month. The problem is not the bar. It is the seats nobody uses. Our rule: measure at day 60, and any seat that cannot show two hours saved a month goes back. Two hours, not one, because in our experience self-reported time savings run optimistic.

This matters in Jersey because most of the firms we speak to here run on Microsoft 365, and the AI Playbook for Jersey notes that the Government of Jersey has itself rolled Copilot out across departments. The seat pitch is landing everywhere. Buy seats for the people whose work is mostly writing, reading and meetings. Build workflows for the processes that run on documents and volume. Different tools, different jobs, and our guide to build, buy or no-code covers the choice.

Budget per case, not per licence

The useful number is cost per unit of work: per document, per case, per client onboarded. Add up model fees, tooling, human review time and a share of the build, divide by volume, and track it monthly. If model fees ever pass 1% of that number, look at the workflow design, because something is sending far more text to the model than it needs to.

What did not change this autumn: the cost of getting your data into shape, wiring the workflow into your systems, and paying people to check the output. The tools are the easy bit. They just got cheaper.

To put your own numbers in, our ROI framework shows the full method. If you want the workflow built with the measurement included, book a review.

Let’s map how your business works

A free 30-minute call with Seb or Mark. Bring the spreadsheet you hate most and we’ll tell you honestly whether a system would fix it.