OpenAI is doubling the usage limits for Work and Codex features across part of its paid plans. The range affected spans from Pro 100 to Business Premium. For teams that use Codex — the tool for AI-assisted code generation and editing — in their day-to-day work, this means noticeably more usable capacity per billing period, without any change to the plan they have booked.

The second part of the announcement is noteworthy: cuts had originally been planned for the Pro 200 plan. Following criticism from users, OpenAI has withdrawn these plans. Anyone working at this tier therefore does not need to revise their usage planning downwards.

Why limits in projects are more than a footnote

In practice, usage limits determine whether an AI tool is suitable as a reliable component of a workflow or merely as occasional assistance. A limit that kicks in in the middle of a refactoring does not just interrupt a session; it shifts tasks, blocks reviews and creates coordination overhead within the team.

That is precisely why doubled quotas are not merely a matter of convenience. They change which work steps can sensibly be delegated to an AI tool in the first place: routine tasks such as adapting recurring components, writing tests or reviewing larger codebases become more plannable when the quota does not have to be calculated tightly.

More efficient models as a second lever

Alongside the raised limits, OpenAI cites more efficient models — named as Terra and Sol — as a factor intended to further increase the capacity actually usable. The mechanism behind this is plausible: if a model requires less compute for the same task, a given quota goes further.

However, there are no robust findings on how strongly this effect translates into concrete development tasks. What remains realistic is therefore this: the doubling of the limits is the hard, verifiable figure, while the efficiency gain of the models is a promise whose effect each team should measure against its own usage.

A signal in B2B competition

The timing is not coincidental. The adjustment comes shortly before the IPO of Anthropic, the direct competitor in the business of AI assistants for enterprises. Anyone wanting to win development teams as a customer segment is long past competing solely on model quality; the question is how much work a plan actually carries per month.

For companies that purchase such tools, this is good news and at the same time an indication of market dynamics: terms in this field are fluid. They can change in favour of users — but the withdrawn cuts also show that adjustments in the other direction remain fundamentally possible.

What this means for budgeting

Those budgeting for AI-assisted development have so far often worked with assumptions based on a few months of experience. The current change is an occasion to review these assumptions — and not only upwards.

  • Reassess plan tiers: If a quota doubles, a lower tier may suffice for a team that previously had to fall back on a higher one.
  • Measure consumption instead of estimating it: Only documented usage over several weeks shows where the actual bottlenecks lie.
  • Limit dependencies: Workflows should be structured so that a reached limit or a change of provider does not bring the project to a standstill.
  • Monitor terms: Plan models in this market are young. An annual review of the framework conditions belongs in the project calendar.

Capacity does not replace methodology

More available quota does not solve structural problems. An AI tool working in a poorly documented codebase will, with double the limit, also produce twice as much material that has to be reviewed. The bottleneck then shifts from the usage limit to the review.

That is why it is worth looking at the accompanying processes: clear conventions in the code, automated tests, a traceable change history and a defined approach to which results may be adopted unchecked and which may not. These foundations determine whether additional capacity translates into speed or into additional rework.

Equally important remains the question of which data may enter an AI tool at all. Larger quotas invite the inclusion of more extensive parts of a codebase — which makes a clean set of rules on confidentiality and trade secrets more important rather than less.

Assessment

For web and software projects, the raising of the limits above all improves plannability: AI assistance can more easily be factored in as a fixed component of a workflow rather than as an add-on that drops away when needed. The real gain, however, only arises where review processes, test coverage and data rules keep pace with the available capacity. Anyone now comparing plans against actual consumption creates the basis for a decision that will still hold after the next adjustment in the market.

Sources