Stop measuring AI by tokens. Measure cost per accepted outcome.

AI is getting cheaper. AI work is not automatically getting cheaper.

That distinction matters now.

For the last couple of years, a lot of the AI conversation has been stuck on model price. Tokens got cheaper. Context windows got bigger. Smaller models got better. Everyone got a bit smug about how much intelligence you could rent for pennies.

Fine. Useful. Not enough.

Because businesses are not buying tokens. They are buying work.

They want the month-end report drafted. The sales follow-up prepared. The campaign brief built. The invoice reminder queued. The website experiment shipped. The spreadsheet checked. The customer research turned into a decision someone can trust.

And once AI moves from a chat box into that kind of workflow, the economics change.

Token price is not the whole story

OpenAI said the quiet part fairly plainly this week in its piece on managing AI investments in the agentic era. Token price has fallen hard. It says price per million tokens dropped 97% from GPT-4 to GPT-5.4, and that GPT-5.6 can complete certain coding-agent tasks with fewer output tokens and less time.

Good news.

Then comes the more important line: token price alone does not show whether AI is creating value. Leaders should look at useful work per dollar.

That is the bit worth paying attention to.

Not because OpenAI has suddenly become a neutral finance department. It is still selling the thing. But the framing tells you where the market is going.

The serious AI conversation is moving from look what the model can do to show me the cost of work I can actually use.

That is a much better conversation.

A cheap model can still make expensive work.

If it fails twice, retries badly, asks for vague clarification, creates a draft nobody trusts, needs a senior person to rewrite half of it, then burns another hour in review, it was not cheap. It was just cheap at the token layer.

The actual cost was hidden in the workflow.

Where businesses get fooled

This is where a lot of businesses will get caught. They will look at an AI dashboard and see usage going up. More prompts. More agents. More connected tools. More scheduled tasks. More people using AI.

That might be good.

It might also be a very modern way to waste time.

Usage is not value. Activity is not value. A folder full of generated drafts is not value. A long-running agent task is not value by default. It becomes valuable when the output is accepted, used and tied to a result the business cares about.

The metric: cost per accepted outcome

So the metric I would use is blunt: cost per accepted outcome.

Not cost per token. Not cost per prompt. Not number of AI tasks completed. Cost per accepted outcome.

What did the workflow produce that a human or system actually accepted?

A resolved support case. A qualified lead list. A reviewed campaign brief. A published page. A merged code change. A corrected invoice run. A sales deck used in a real meeting. A CRO test shipped with a clean before-and-after measure.

Then count the full cost of getting there: model and tool usage, failed attempts, retries, latency, human review time, corrections, approval time, risk checks, and the cleanup when it gets something wrong.

Suddenly the economics look different.

Three places to test it

Take the AI tools assessment note from yesterday’s Startup Ideas capture. The useful bit was not “use Claude to make a report”. Plenty of people can generate a report. The useful business is paid diagnosis: a discovery call, a transcript, pain extraction, tool recommendations, a short report, human QA, a review call, then implementation.

The accepted outcome is not the AI-generated report draft.

It is a client saying, yes, that is our pain, those are the first fixes, and we trust you to help us implement them.

That is the outcome worth measuring. How long did it take to produce a report good enough to earn that conversation. How much human QA was needed. How many recommendations were actually implemented. How many assessments turned into paid build work.

That is a business metric. The token bill is trivia by comparison.

Same with the website-as-revenue-agent idea from the Marketing School note. A site that runs weekly CRO experiments with an agent in the loop sounds clever. But the accepted outcome is not “the agent suggested five tests”.

The accepted outcome is a test shipped, measured and learned from.

If the agent produces twenty variant ideas and the operator only trusts one, the cost is not the price of generating twenty ideas. It is the total time and spend needed to get one useful test live.

That is what should go on the scoreboard.

Coding agents make this even clearer. The GitNexus note is a good example because it treats agent code edits like work with a blast radius. Before an agent changes a symbol, it should know what calls it. After it changes the code, it should detect affected flows. The accepted outcome is not “agent edited files”. That can be a disaster with syntax highlighting.

The accepted outcome is a change that passes tests, survives review and does not break the parts of the system nobody remembered to mention.

Again: cost per accepted outcome.

Why “AI is cheap now” is a half-truth

This is also why “AI is cheap now” is a dangerous half-truth.

Yes, model costs can fall. But as agents become more capable, we ask them to do longer, messier work. They use more tools. They touch more systems. They run on schedules. They pull in more context. They require governance, permissions, logging and review. They move from answering into operating.

That can be worth it. Often it will be.

But only if the workflow is measured like work.

The operator checklist

A sensible AI operating sheet for a workflow should answer a few boring questions before anyone scales it. What accepted outcome are we paying for. What is the current human cost of producing it. What does the AI workflow cost end to end, including review. What quality bar does the output have to meet. What happens when it fails. Who approves the output. What should never be automated. At what cost does the workflow stop being worth running.

That last one matters. A lot.

Because agent work has a nasty habit of feeling productive while it wanders. A human operator will often stop when the work is obviously going nowhere. An agent may keep politely trying: calling tools, rewriting, searching, retrying, and producing more stuff for someone to inspect.

That is not intelligence. That is an invoice with a friendly tone.

The answer is not to ban it. The answer is to put a cost boundary round the work.

For a campaign brief, maybe the agent gets one source pass, one draft pass and one revision pass before a human takes over.

For sales research, maybe it gets a fixed source list and must return a confidence score with every claim.

For code edits, maybe it cannot touch high-risk flows without impact analysis and a named reviewer.

For website CRO, maybe it can suggest tests every week, but only one approved test ships and the metric is conversion lift, not idea volume.

This is where the post-agency work sits

The old agency sold output: pages, ads, posts, decks, reports.

The lazy AI agency sells speed: “we can make more of that, faster.”

The useful AI operating partner sells a measured workflow: source set, cadence, quality bar, approval point, cost cap, evidence trail and accepted outcome.

That is less glamorous. Good. Glamour is usually where the margin leaks out.

If I were advising a business owner on AI spend now, I would not start with model choice. I would start with one workflow and a scoreboard.

Pick something that happens every week: a sales follow-up review, a content brief, a quote pack, a customer insight report, a website test, a finance check, a support triage pass.

Write down what “accepted” means. Run it manually once. Run it with AI once. Count the full cost both ways. Include the review time. Include the rework. Include the bits nobody wants to admit, like the senior person cleaning up the clever draft.

Then decide.

If the AI version is faster, cheaper or better at the quality bar, scale it carefully.

If it produces more noise, kill it.

That is not anti-AI. That is adult supervision.

The companies that win with agents will not be the ones with the most AI usage. They will be the ones that know which machine work is worth accepting.

---

Pull quotes

• “Businesses are not buying tokens. They are buying work.”
• “A cheap model can still make expensive work.”
• “Usage is not value. Activity is not value. A folder full of generated drafts is not value.”
• “The accepted outcome is not ‘agent edited files’. That can be a disaster with syntax highlighting.”
• “That is not intelligence. That is an invoice with a friendly tone.”

Short LinkedIn / X version

Token prices are falling.

That does not mean your AI work is automatically getting cheaper.

Businesses do not buy tokens. They buy work: a campaign brief they can use, a sales follow-up that gets sent, a website test that ships, a code change that passes review, a report that changes a decision.

So the useful metric is cost per accepted outcome.

Count the full cost: model spend, tool calls, retries, review time, corrections, approvals and cleanup.

A cheap model can still make expensive work if it creates drafts nobody trusts.

My rule: pick one weekly workflow, define what “accepted” means, run it manually, run it with AI, then compare the real cost.

Scale what proves itself. Kill the clever noise.

Notes and caveats

• OpenAI’s 97% token-price drop and GPT-5.6 efficiency claims are OpenAI’s own framing. Useful signal, but vendor-supplied.
• Lower token prices are not irrelevant. They matter. They just do not prove workflow ROI on their own.
• The Google Ads Advisor source is from April 2026, used only as background on the broader shift toward agentic commercial operations.
• The Startup Ideas assessment economics are podcast-source claims. Good as an offer pattern, not as audited proof.
• If publishing, swap “Foundry” or “post-agency” references depending on where this runs.