Within two days, two announcements have appeared that raise the same question for development teams: how stable is the foundation on which AI-supported development now runs?OpenAI has terminated Cursor AI coding tool's access to its models.Anthropic is changing the weekly usage limits of Claude Code.Neither is a technical failure, but decisions by providers – and that is precisely what makes them relevant to planning.

Case 1: Cursor loses OpenAI access

Cursor is a development tool that generates code suggestions and changes via large language models.Cursor does not operate these models itself but obtains them from providers.Following the acquisition of Cursor by SpaceX, OpenAI terminated the contract.The stated reason is that Elon Musk's companies had repeatedly breached contracts.

Cursor co-founder Michael Truell plays down the consequences: OpenAI accounts for only five percent of AI traffic.Whether this assessment holds true in users' everyday work cannot be inferred from the announcement.Noteworthy is another point: the trigger for the loss was not a price dispute or a capacity problem, but a change of ownership and the resulting conflicts at contract level.Such causes can hardly be foreseen by technical means.

Case 2: Claude Code with newly adjusted weekly limits

At Anthropic, it is not about access but about quantity.Claude Code is Anthropic's command-line tool that lets users hand off code tasks directly to a model.Usage is subject to weekly limits, i.e. a capped capacity per billing period.

According to the announcement, a temporary 50 percent increase of this capacity ends on 14 September.It is replaced by a permanent increase of 25 percent.Compared with the current state, this means less available capacity; compared with the baseline before the temporary increase, more.In return, Anthropic promises more control and transparency over consumption.

For teams that have grown used to the higher level, this is a noticeable change to their working framework – and on a fixed date.Anyone who has factored capacity into sprints or release planning must recalculate.

What the two cases have in common

The causes differ, the pattern is the same: the performance of one's own toolchain depends on third-party decisions that can be made at short notice.This creates two different levels of dependency that are often conflated in practice.

  • Model dependency: A particular model is suddenly no longer available.But prompts, conventions and established working practices are aligned with its behaviour.
  • Provider dependency: The tool itself – the interface, the integration into the development environment, the billing – lies with a provider whose business relationships can change.
  • Capacity dependency: It is not access that is missing, but the usable quantity that falls.This particularly affects teams that have firmly embedded AI support into their workflows.

Practical consequences for the toolchain

No recommendations for specific products can be derived from the two announcements.But questions that are worth asking before the next subscription renewal.

  1. Which models do we actually use – and for what? Anyone with a simple list of which model is used in which work step can switch in a targeted way during an outage instead of improvising.
  2. Is switching technically provided for? Tools that support multiple providers reduce the risk of a single contract ending.The Cursor case shows that diversity on the model side helps, even when one share drops away.
  3. How much capacity do we realistically need? Consumption data is the prerequisite for budget planning.Announced transparency features are only useful if someone evaluates them regularly.
  4. What happens when the limit is reached? There needs to be a defined state for the case that no capacity is available mid-week – including the question of which tasks then take priority.
  5. Does the work remain possible without AI support? If reviews, tests and documentation run exclusively through one tool, a change of provider becomes a project risk.

Cost planning becomes an ongoing task

Both cases show that the cost and capacity side of AI tools cannot be calculated once.Temporary increases can expire, contracts between providers can end.Anyone using AI support productively should treat these items like other operating costs: with a fixed date for review, with documented consumption and with an assessment of which alternative applies if needed.

It also makes sense to sort dependencies by criticality.A tool that supports individual developers in writing code is more replaceable than an automation that is integrated into build or deployment processes.The deeper the integration, the more carefully the exit path should be described.

Assessment

For web and software projects, this means: AI tools are now part of the infrastructure and should receive the same attention as hosting, databases or interfaces to third-party systems.Anyone who keeps models interchangeable, makes consumption measurable and does not unlearn how to operate without AI support remains able to act when a provider makes different decisions at short notice.

The two reports are less of an alarm signal than an occasion to take a deliberate look through your own toolchain.

Sources