Model dependency: what a model swap or a price change does to the plan

Model dependency risk is the exposure a company carries because the model in its product belongs to a supplier who can retire it, reprice it, change what it does, or ship a product that competes with the company's own. Four events, four costs. A company that has named all four and can say what each would cost has managed the risk. A company whose plan assumes one model at one price for three years has a supplier risk nobody priced.

This page takes each event in turn and ends with the test to run in diligence. The margin arithmetic is on rebuilding inference gross margin from invoices; the classification of how much of the product is the model is on AI feature or AI wrapper.

What do the providers promise about retirement?

The providers promise notice, and the notice is short enough to matter. Reveneau's diligence asks every AI target for its model inventory with a retirement date and a migration plan against each entry, because both major providers publish the schedule and both have used it.

OpenAI's deprecations page states that generally available models get at least 6 months' notice before retirement, specialised variants at least 3 months, and preview models may be retired on much shorter notice, giving 2 weeks as an example. Its recent entries include the Assistants API, announced 26 August 2025 for shutdown on 26 August 2026, and a specialised model, gpt-5.4-cyber, announced 11 September 2026 for shutdown on 1 October 2026 [1]. Anthropic's page states that customers with active deployments get at least 60 days' notice before a publicly released model is retired. Its history shows Claude Sonnet 4 and Claude Opus 4 deprecated on 14 April 2026 and retired on 15 June 2026, Claude Opus 4.1 deprecated on 5 June 2026 and retired on 5 August 2026, and Claude Sonnet 3.7 deprecated on 28 October 2025 and retired on 19 February 2026; each entry names a recommended replacement, and the page states that requests to retired models fail [2].

So a company that pinned a model version in 2025 has, in all likelihood, already been through one forced migration, and the diligence question is what it looked like. Anthropic's own guidance on the same page is to test applications with the replacement model well before the retirement date [2]. A company with an eval suite does that in an afternoon. A company without one finds out in production.

Event	Provider rule (published)	What it does to the plan
Retirement of a GA model	OpenAI: at least 6 months' notice. Anthropic: at least 60 days' notice	Forced migration; cost is the eval re-run plus any prompt rework
Retirement of a preview or specialised model	OpenAI: 3 months for specialised variants, previews as short as 2 weeks	Same, on a schedule the company does not control
Price change	No notice rule stated on either pricing page	Margin moves in the same month, per the invoice
Capability change on a new version	Replacement model named by the provider	Eval scores move; the suite says which way
Provider ships the product	None	Competition from the supplier, on the supplier's price list

What does a price change do to the plan?

A price change lands on the margin in the month it takes effect, in whichever direction it goes, and the plan should say what it assumes. The direction over the last few years has been down for a fixed level of capability. Stanford's AI Index 2025 reports that the cost of querying a model at GPT-3.5 level fell from $20.00 per million tokens in November 2022 to $0.07 by October 2024, a more than 280-fold reduction, and that inference prices have fallen between 9 and 900 times a year depending on the task [3]. A company that has been in business through that period has had its cost of goods fall without doing anything, and its margin history makes its margin management look better than it is.

The same direction is not guaranteed for a specific product. Anthropic's pricing page records that the introductory price for Claude Sonnet 5, $2 per million input tokens and $10 output, announced as running to 31 August 2026 with a scheduled rise to $3 and $15 on 1 September, was instead made the standard price; the rise did not happen [4]. That is a price decision the company on the other side of the invoice did not make and could not have planned for in either direction. The same page notes that the tokenizer used by Claude 4.7 and later produces more tokens for the same text, giving 30 percent as the figure, so a model upgrade can raise the bill for identical usage [4]. A plan that assumes prices only fall is a plan; a plan that shows the margin at list price plus 30 percent is a stress test.

What does a model swap cost?

A model swap costs an eval run plus whatever the eval run finds, and for a company without a suite it costs whatever production finds. The eval suite is what makes the swap a configuration change rather than a rebuild; the eval suite as a diligence artefact is the page on reading one.

The buyer side shows what is possible. a16z's June 2025 survey of 100 CIOs found 37 percent running five or more models in production, up from 29 percent the year before, and reported that differentiation by use case is the main reason for buying from several vendors; the same survey reported that switching costs are rising as agentic workflows create dependencies that take engineering effort to reproduce on another model [5]. Menlo Ventures' December 2025 survey put enterprise LLM API spend at 40 percent Anthropic, 27 percent OpenAI and 21 percent Google [6], and ICONIQ's July 2026 survey found companies running 3.3 models on average, with nearly half using two or more model types [7]. Buyers switch, and buyers with complex workflows find switching expensive. A startup's swap cost sits somewhere on that range, and the eval suite is the only artefact that says where.

What does provider competition do?

Provider competition replaces the product with a feature on the supplier's price list. Foundation Capital's September 2026 analysis puts it in one sentence: the model provider that powers you can also turn around and compete with you. It argues that providers see what works across everything built on their platforms, names OpenAI's agent products and Anthropic's Claude Code as examples of providers shipping into application territory, and describes the "Windsurf saga" as a case of preferential treatment followed by competition [8]. Its survival advice, high-exception workflows, subjective success criteria and fragmented legacy systems, is a list of what providers will not build, and it is the same list as the feature side of the wrapper test.

For diligence, the question is what the product would be worth on the day the provider ships the core of it. If the answer is the same, because the company owns the workflow, the integrations and the customer's data loop, the dependency is a supplier relationship. If the answer is nothing, the dependency is the business.

How do you test the dependency in diligence?

Test the dependency by making the company do, in the diligence session, what a retirement would force it to do.

1. Ask for the model inventory. Every model and version the product calls, the date each was pinned, the provider's published retirement rule for it, and the named migration target.
2. Ask what happened on the last retirement. Which model, how much notice, how long the migration took, what broke. Anthropic's list of 2025 and 2026 retirements means every company on its models has a story [2].
3. Ask for the swap to be run live. The eval suite on the second-choice model, in the session. Time it and read the scores. An afternoon is a managed dependency. "We would need to rework the prompts" is a rebuild.
4. Re-run the margin at stress prices. The per-customer table at list price plus 30 percent, and at the second-choice model's price. Read the bottom rows.
5. Ask what the product is on the day the provider ships it. Take the answer to the wrapper test and write it down.
6. Check the terms. Which account type, which terms, which retention, for each provider, because a migration to a new provider is also a migration to new terms.

What goes in the report?

The report names the four events, the company's exposure to each, the measured cost of the live swap, and the margin at stress prices. The consequence for the deal is usually one of two sentences. "Dependency managed: inventory current, last migration took two days, live swap passed the suite, margin holds at plus 30 percent." Or: "Dependency unmanaged: single pinned model, no suite, swap would be a rebuild, margin negative at the second-choice price." Data moat claims covers the one defence that is supposed to survive all four events, and the AI startup due diligence guide puts model dependency beside margin and moat in the money read.