Two of Anthropic's largest corporate customers are significantly reducing their use of Claude: Microsoft cut the monthly per-head budget in its cloud division from 100,000 to 10,000 dollars.Meta halved the number of employees with access to Claude Code to 30,000.Both corporations are instead increasingly relying on their own AI tools.
For Anthropic, the close ties to a few major customers thus become a strategic risk.For all other companies, the development is above all an indication of how quickly toolchains in software development can shift – and how unreliable the assumption is that an assistant chosen today will remain in place for years to come.
What is interesting about these figures
The budget cut at Microsoft is not a small correction step, but a reduction by an order of magnitude.That points less to dissatisfaction with a single product than to a changed procurement logic: where a corporation operates its own models and its own development tools, every external purchase becomes a question of internal cost allocation.
At Meta, the same pattern shows in the number of users.It is not that the tool disappears, but the circle of those permitted to use it is drawn more narrowly.In both cases, it remains open how much weight professional reasons, cost pressure or strategic independence tipped the scales.The direction is nevertheless clear: away from broad licensing of an external assistant, towards internal alternatives.
The situation in mid-sized businesses is different – the risk is not
Mid-sized companies and agencies cannot train their own language models or build a complete internal development platform.They depend on external providers.That is precisely why a shift in the provider market affects them differently, but no less harshly.
If a provider adjusts prices, changes quotas, discontinues a model or restricts availability in certain regions, that cannot be absorbed by an internal replacement product.The only option is to switch – and the deeper a single tool has dug itself into daily work, the more expensive that is.
How to recognise dependency in everyday development
Technical lock-in rarely arises from a conscious decision, but creeps in.Typical signs:
- Prompts and instructions are scattered across provider-specific configuration files instead of being in a central, documented place.
- Automations in the build pipeline – that is, in the automated processes for tests and delivery – call a vendor-specific interface directly, without an intermediate layer.
- Code review processes assume that a particular assistant is available and have no defined fallback path.
- Knowledge of good working practices sits in individual people's heads, not in verifiable guidelines.
None of these points is dramatic in itself.Taken together, they mean that switching providers is no longer a configuration matter, but a project.
Working model-independently: practical starting points
Independence does not mean doing without AI support or constantly paying for several subscriptions in parallel.It means keeping switching costs low.
- Introduce an abstraction layer: Calls to language models go through a slim interface of your own in the project, not directly from the application code.A model change then affects one place instead of thirty.
- Version prompts: System instructions, coding guidelines and project context belong in the repository, not in a tool menu.That way they are transferable and traceable.
- Make results measurable: Anyone who collects test coverage, error rates and runtimes can evaluate an alternative assistant objectively instead of deciding by gut feeling.
- Keep quality assurance independent: Automated tests, static code analysis and binding reviews must work regardless of which model supplied the draft.
- Document data flows: Which source code, which customer data goes to which service?This documentation is necessary anyway and makes a switch considerably easier.
Plan contracts and budgets realistically
The development at the two major corporations also shows that budgets for AI tools should be revisable at short notice.Long terms with fixed user numbers fit poorly with a market in which performance and pricing models change every six months.
It makes more sense to initially limit use to clearly defined use cases, to prove the benefit there and to align the licence volume with it.An annual review date at which tools, costs and alternatives are compared soberly replaces the principle of hope.
Assessment for web and software projects
For web and software projects, this above all means: The quality of a project must not depend on the specific AI assistant, but on architecture, tests and traceable processes.Anyone who cleanly separates prompts, guidelines and interfaces from the respective provider can deploy a better or cheaper model as soon as it is available – without the project coming to a standstill.What matters is therefore not the choice of a tool, but the ability to be able to revise that choice at any time.