Skip to content
Blentera

Resources, AI strategy

How to avoid lock-in to a single AI model or vendor

Providers retire AI models on fixed schedules and release new ones every few months. Companies that build agents around one model end up rebuilding when it changes. This guide shows where lock-in really sits and how to keep the freedom to switch.

Updated 20 November 2026. 9 minute read.

In short

  • Models are retired on a schedule, sometimes with only weeks or months of notice. Plan for every model you use to be replaced.
  • Lock-in rarely comes from the model alone. It comes from prompts tuned to one model, memory stored with one vendor, custom integrations, fine-tunes, proprietary formats and contract terms.
  • Keep company knowledge, rules and history in systems you control, outside the model, so a new model can pick them up.
  • Use provider-neutral access, such as your own accounts with several providers or a router, and test models per task before switching.
  • Write an exit plan with export formats and notice periods. The EU Data Act's switching rules cover cloud services and remove switching charges from 12 January 2027.

Why models change so fast

AI providers manage models as products with a lifecycle. OpenAI and Anthropic both publish deprecation pages. These list models with a shutdown or retirement date and a recommended replacement. Once a model is retired, requests to it fail.

Notice can be short. Anthropic commits to at least 60 days' notice before retiring publicly released models. OpenAI's deprecation policy says preview models may be retired at much shorter notice. A concrete case: OpenAI released GPT-4.5 as a research preview in February 2025, announced its deprecation in April 2025 and shut it down in the API on 14 July 2025. It named GPT-4.1 as the replacement.

Fine-tuned models follow their base model. On 4 January 2024, OpenAI shut down its older GPT-3 base models together with fine-tunes built on them. Customers with those fine-tunes had to train new ones on a newer base model.

Treat every model as temporary. The question is not whether you will switch, but how much work each switch costs.

What actually locks you in

Changing the model name in an API call takes seconds. The cost of switching sits in everything built around it:

  • Prompts. Instructions tuned over months to one model's habits often behave differently on another. Without tests, you cannot tell what broke.
  • Memory. Conversation history, notes and embeddings kept inside one vendor's assistant or thread store may not export in a usable form.
  • Integrations. Tool connections written against one provider's function-calling format need rework for another.
  • Fine-tunes. A fine-tuned model lives on the provider's platform and shares its base model's lifecycle. Your training data is portable. The tuned model usually is not.
  • Data formats. Logs, files and agent state saved in proprietary formats are hard to move.
  • Contracts. Committed spend, long notice periods, exit fees and unclear data return terms all raise the cost of leaving.

Open standards reduce the integration part. The Model Context Protocol, a standard for connecting AI applications to tools and data, was donated by Anthropic to the Agentic AI Foundation under the Linux Foundation in December 2025. OpenAI and Block were co-founders of that foundation. Tools exposed this way can be reused across clients that support the protocol.

Keep company knowledge and rules outside the model

The most valuable part of an agent setup is what it knows about your company and the rules it follows. Keep both in places you own:

  • Company knowledge in a memory or document store you control, in plain text or open formats, retrieved and passed to whichever model runs the task.
  • Rules and permissions in the platform, not in the prompt. Which agent may send email, spend money or change records should be enforced by the system, whatever model is behind it.
  • Approval steps and budgets attached to roles and projects, so they survive a model change.
  • A full activity log in your own systems, so you can compare old and new models on real work.

This also helps with data protection. Under the GDPR, individuals can ask you to correct or erase their data. That is far easier in a store you control than in a vendor's memory feature or a fine-tuned model. Note that the GDPR's portability right (Art. 20) belongs to individuals, not to companies. For company data, portability depends on your contracts and the Data Act.

Use provider-neutral access

Keep at least two providers ready for important work. There are two common ways to do it.

Your own accounts with several providers

Hold direct accounts with more than one provider. You get each provider's own contract, data terms and invoices, and you can move a workload by changing a setting rather than signing a new deal.

A router

A router gives one API in front of many models. OpenRouter, for example, offers an OpenAI-compatible endpoint, so switching is mainly a change of base URL, key and model name. Its documentation describes model fallbacks, where a list of models is tried in order if the first is down or rate-limited, and provider routing settings that control which providers may serve a request. Check a router's own terms and data handling, because it becomes another party in the chain.

Whichever route you take, keep your own code and prompts free of one provider's special features unless the gain is worth the switching cost. Wrap any such feature so it can be replaced.

Evaluate models per task

No single model is best at everything, and rankings change with each release. Pick models per task, not per company:

  1. List the agent tasks that matter, such as triaging support emails, drafting reports or reviewing code.
  2. Build a small test set for each from real, anonymised examples, with what a good answer looks like.
  3. Run candidate models on the same set. Score quality, cost per task and speed.
  4. Have an employee who knows the work review a sample of the outputs.
  5. Re-run the tests when a provider announces a deprecation or a new model, and before any switch.

Those test sets are themselves an asset that reduces lock-in. With them, a model change becomes a measured decision instead of a leap.

Exit planning and data portability

Write the exit plan before you need it. For each AI vendor and platform, record what data you hold there, how to export it, in which format, and how long the export takes.

The EU Data Act's switching rules

The Data Act (Regulation (EU) 2023/2854) has applied since 12 September 2025. Its Chapter VI covers switching between data processing services, meaning cloud and edge services. The main rules are:

  • the maximum notice period to start a switch is two months
  • the transitional period for the switch is at most 30 calendar days, unless technically unfeasible. The provider must then say so within 14 working days, and the period may be extended to no more than seven months
  • without common standards, exportable data must be provided in a structured, commonly used and machine-readable format
  • until 12 January 2027, providers may charge reduced switching fees that do not exceed their direct costs. From that date, switching charges are no longer allowed

Whether a given AI service is a data processing service under the Act depends on how it is provided, so check with your legal adviser. In November 2025, the Commission proposed amendments to the Data Act that would ease some switching duties for custom-made services and smaller providers. Check whether they have been adopted before you rely on the exact rules.

The checklist

  1. List every model and provider your agents use, with its deprecation status and the provider's notice policy.
  2. Keep company knowledge, rules and history in stores you control, in open formats.
  3. Enforce permissions, approvals and budgets in the platform, not in prompts.
  4. Hold accounts with at least two providers, or use a router, for important workloads.
  5. Avoid provider-specific features unless they are wrapped and replaceable.
  6. Prefer open standards such as the Model Context Protocol for tool connections.
  7. Keep fine-tuning data so you can retrain on another base model.
  8. Maintain test sets per task. Re-run them on every model change.
  9. Read contracts for committed spend, notice periods, exit fees and data return.
  10. Write and test an exit plan: what to export, in which format and how long it takes.

How Blentera helps

Agents work from one company memory that stays when you switch AI models. Roles, approval rules, budgets and the activity log sit in the platform, not in a model. You connect your own provider accounts: OpenAI, Anthropic and xAI directly, or Gemini, Mistral and others through OpenRouter. You pay providers directly, with no markup. The platform is hosted in the EU.

Questions

Sources

  1. OpenAI API: deprecations
  2. OpenAI API: changelog
  3. OpenAI: Introducing GPT-4.1 in the API
  4. OpenAI: GPT-4 API general availability and deprecation of older models in the Completions API
  5. Anthropic: model deprecations
  6. Anthropic: donating the Model Context Protocol and establishing the Agentic AI Foundation
  7. Linux Foundation: announcement of the Agentic AI Foundation
  8. OpenRouter: OpenAI compatibility
  9. OpenRouter: model fallbacks
  10. OpenRouter: provider routing
  11. Regulation (EU) 2016/679 (GDPR), EUR-Lex
  12. Regulation (EU) 2023/2854 (Data Act), EUR-Lex
  13. European Commission: Data Act explained
  14. European Commission: Digital Omnibus proposal amending the Data Act, COM(2025) 837

This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.

Get started

Put your first agent to work in minutes.

Three steps, and your first agent is working for your company.

  1. Step 1

    Create an account

    Sign up with your work email and invite your team. No sales call needed.

  2. Step 2

    Add your first agent

    Create an agent, give it a role, and connect your own AI key and the apps it may use.

  3. Step 3

    Give it a task

    Ask in Blentera or your chat app, and follow the work until it is done.