AI Vendor Lock-In: 7 Signs One Model Owns Your Workflow

You have AI vendor lock-in when one model provider holds your workflow so tightly that you cannot switch, or credibly threaten to switch, when the price, the terms or the model itself changes. It rarely starts as a decision. A team writes a prompt that works, wires in one provider's tool features, tunes around one model's habits and ships. The cost shows up months later, when the provider raises a price or retires a model, and the only choices left are pay or rebuild.

Why lock-in is a pricing problem

Google announced Gemini 4 Argon at an introductory $2 per million input tokens and $10 per million output tokens, and says the price rises to $4 and $20, without saying when. A workflow priced at the launch rate will cost twice as much per token after that.

The same logic applies to every term a vendor can change, from a cache discount to a retirement date. Your leverage on each one is the cost of leaving. Our comparison of Argon and Claude Sonnet 5.5 shows how much one price row can hide.

Seven warning signs

1. The workflow lives in one provider's format. Your business logic sits inside one provider's prompt format or agent framework, so moving means rewriting.
2. Nobody has a list of model IDs. No inventory of the exact model versions you call, such as claude-sonnet-4-5-20250929, or of which workflows use them. You cannot price a change you cannot see.
3. You cannot run the same task elsewhere. Sending one task to another model means rewriting the integration, not changing a setting.
4. You have no golden test set. That is a fixed batch of real tasks with known good answers. Without one, you cannot prove an alternative is good enough, so the threat to switch is empty.
5. Tools depend on vendor-only functions. Your tools and data access rely on one vendor's features, with no documented interface another model could use.
6. Nobody watches the vendor. You do not track pricing, deprecations, availability or policy changes, so you learn an introductory price ended from the invoice.
7. There is no plan B. No fallback model, no manual process and no migration playbook. When the model retires, the workflow stops.

The signs escalate. The first five make leaving slow and expensive. The sixth means you will not see the change coming. The seventh turns a price change or a retirement into an outage.

Five terms to settle before you sign

When prices are moving, the terms around the price matter as much as the price. The table pairs each question with what the two big labs have published as of October 2026.

A cache discount rewards prompts built the way one provider's cache works, so part of your savings may be tied to one vendor. Retirement terms have a catch too. Anthropic's deprecation page says its dates apply to its own platforms, while partner platforms such as Amazon Bedrock and Google Cloud "set their own retirement schedules." If you reach a model through a cloud partner, the date that matters is the partner's.

How to reduce the dependency

Start with visibility: map your models, prompts, tools, data and owners. For Claude, Anthropic's deprecations page points to the Usage page in Claude Console, whose Export gives a CSV of usage by API key and model. Then work through five changes:

• Centralize routing. One place in your code decides which model gets a request, so a swap is a setting. Our post on model-agnostic AI architecture covers how to build that layer.
• Move policies out of prompts. Approval rules and data boundaries live in your system.
• Standardize output formats. Every model must return the same structure, so the code after it does not change.
• Keep your evaluation data. The golden test set and past results let you judge a replacement in days.
• Build a small fallback path for the critical workflows, even if it is slower or more manual.

Do not try to rebuild every workflow at once. Begin with the ones that touch customers, revenue, regulated data, production systems or high volume. Our posts on model deprecation and LLM API cost control cover baseline tests and routing. A routing service can turn a swap into a configuration change, as our post on Stripe's OpenRouter acquisition describes, but it is one more vendor in the chain.

When staying put is the cheaper choice

Portability costs money. A fallback model needs its own test runs, its own prompt tuning and an account kept in working order, and a golden test set needs upkeep as the work changes. For an internal summarizer that a person reads before acting, all of that can cost more than the lock-in it prevents. There, a price rise is annoying and a retirement means a week of rework.

The money is better spent where a forced switch would stop revenue or reach customers. Most workflows sit between those two cases, and where to draw the line is a judgment about what an outage would cost, made one workflow at a time.

Pairs well with

At renewal, the AI Vendor Reliability & Spend-Justification Scorecard ($79) scores each AI vendor on reliability and value and returns RENEW, RENEGOTIATE or DO NOT RENEW. The Vendor Auto-Renewal & Contract-Trap Tripwire ($49) computes each contract's notice deadline from the terms you confirm and reads it ACT NOW, WINDOW OPEN or LOCKED IN. When you test a replacement model, the Prompt Regression Lab ($89) A/B compares its outputs against a saved baseline and gives a ship, hold or regressed verdict.