Are model price drops a reason to compete on workflow rather than inference cost?

Model price drops should make your workflow better, not turn your company into a token-price spreadsheet.

Stanford's 2025 AI Index reported that the cost of querying a model at roughly GPT-3.5-level MMLU performance fell from $20 per million output tokens in November 2022 to $0.07 by October 2024. That is the kind of change that makes founders redraw a product plan.

But cheaper inference is not the same thing as cheaper work.

The current price snapshot

Prices move, so this is a dated reference, not a permanent rate card. I checked the official pages on September 19, 2026.

Provider/model reference	Input	Output	Simple 1M input + 200K output example
Anthropic Claude Sonnet 4, standard API	$3 / 1M tokens	$15 / 1M tokens	$6.00 token cost
Google Gemini 3.8 Flash, standard paid tier	$0.75 / 1M tokens through December 31, 2026	$3.75 / 1M tokens through December 31, 2026	$1.50 token cost
Batch variants	Usually discounted	Usually discounted	Useful for work that can wait

The Anthropic numbers come from its current model-pricing PDF. The Google numbers come from the current Gemini pricing page. OpenAI's live pricing page is the right place to refresh its current list rather than copying a number into a post that will age badly.

The example is only arithmetic. It does not say these models are equally capable, equally fast, or interchangeable. It also ignores the expensive bits that appear after the first API call: retrieval, tool calls, retries, long context, storage, human review, support, security, and the sales process required to get access to real data.

Four ways to spend the savings

1. Make the existing workflow cheaper

This is the obvious choice. It is also the least defensible one if the customer can switch providers without noticing.

If your product is a thin wrapper around one model call, a price cut is good for the customer and bad for your margin. You can respond by lowering price, but now your advantage is a rate card that the model provider controls.

That can still be a good business. It is just a different business from a workflow product.

2. Buy more reliability

Use the savings to run a second pass on uncertain outputs, retrieve the missing document, call a specialist tool, or send an exception to a human. The customer does not care that your input tokens got cheaper. They care that the work gets finished without a correction loop eating the afternoon.

OpenAI's Basis case study describes this direction. Basis routes accounting tasks to different models based on task, complexity, latency, and input type, then exposes assumptions, sources, and decision logic for review. The page is a vendor-selected customer story, not an independent test. The product lesson is still useful: model choice belongs inside a workflow decision.

3. Expand the job, not the prompt

If a cheaper model lets you handle more steps, the product may become more valuable. A reconciliation tool might also gather source files, explain exceptions, prepare a review queue, and record what the accountant changed.

That is where token savings become workflow competition. The extra value is not “we generated 40% more text.” It is “the customer now finishes the task with fewer handoffs.”

Measure the completed unit. A processed claim, reviewed journal entry, accepted support resolution, or approved research packet is more useful than a token count.

4. Price around the outcome

If inference is falling, cost-plus pricing gets less interesting. That does not mean every agent should charge per outcome. Outcomes can be hard to attribute, seasonal, or politically owned by someone else inside the customer.

It does mean you should know what the buyer is paying for. Is it a lower review burden? More completed work per employee? Faster turnaround? Fewer escalations? A token-based plan is easy to explain, but it makes your economics visible while hiding the customer's value.

The founder decision

I would make the cost decision in this order:

1. Measure the cost of one completed customer outcome, including retries and human review.
2. Use the cheapest model that clears the quality bar for routine steps.
3. Reserve expensive reasoning for ambiguity, exceptions, and high-cost mistakes.
4. Turn the saved capacity into a larger completed workflow before lowering price.
5. Refresh the official price table whenever a provider changes a model or batch policy.

Price drops are a tailwind. They are not a moat. If the customer can feel the savings only because your bill is lower, the provider owns the advantage. If the savings let you finish a harder, more trusted workflow, you might own something.