OpenAI is changing how ChatGPT is used: the previously separate Chat and Work modes will be merged into a single interface. However, the change behind the interface is more noticeable than the interface itself. From December, the Plus, Pro and Business plans will draw on a shared, token-based allowance – for both normal ChatGPT operation and Codex, the tool for AI-assisted software development.
Tokens are the billing unit of language models: small text or code fragments into which inputs and responses are broken down. Anyone who has previously thought in categories such as „number of messages per time window“ must now switch to a consumption logic based on the actual amount of text processed.
What exactly is changing
The key points of the change can be summarised briefly:
- Chat and Work no longer appear as separate modes, but in a single interface.
- Plus, Pro and Business plans receive a shared allowance from which both ChatGPT and Codex draw.
- More complex models consume more of this allowance than simpler ones.
- Simple instant chats are not counted against the allowance.
This creates a tiering that did not previously exist in this form: not every request costs the same. A short follow-up question in everyday use carries different weight than a longer-running Codex run that reads a larger codebase, suggests changes and writes results back.
Why this matters for development teams
Codex sessions are typically token-intensive. Context from multiple files, test output, error messages and iterative correction loops quickly add up. If this consumption sits in the same pot as everyday ChatGPT use, two very different ways of working compete for the same budget.
In practice, this means: a team in which the same accounts are used both for text work – concepts, summaries, research – and for development tasks can exhaust the allowance much faster than previously calculated. The article describing the change points out that intensive use in the cheaper plans is likely to lead to quality losses, because less sophisticated models will then have to be used instead.
This assessment is a forecast, not a commitment from OpenAI. But it describes a plausible mechanism: if more capable models cost more tokens and the budget is limited, there is an incentive to choose the more economical model when in doubt.
Cost planning becomes consumption planning
Previously, the calculation was simple for many companies: a fixed monthly fee per workstation. With a shared token-based allowance, the question shifts from „How many licences do we need?“ to „How much consumption does which activity generate?“.
This requires a fundamentally different approach to budgeting:
- Monitor consumption: Before adjusting plans, it is worth having a phase in which actual consumption per role and task is recorded.
- Separate usage: Where possible, development-related and general usage should not be mixed uncontrollably in the same account, so that bottlenecks remain attributable.
- Steer model choice deliberately: Not every task needs the strongest model. Clear internal guidelines on when which model makes sense save budget without a noticeable loss of quality.
- Keep alternatives open: Dependence on a single provider becomes more expensive when its billing logic changes. Those who design interfaces and workflows so that a switch remains possible stay able to act.
Impact on tool selection
For teams that have firmly integrated Codex into their everyday development work, the question of the right plan tier arises anew. Those who previously managed with a cheap plan because usage was spread across several tools could hit a limit due to the consolidation.
At the same time, a more differentiated view of one's own workflows emerges. The distinction between counted and non-counted requests – simple instant chats do not count against the allowance according to the current presentation – makes it sensible to deliberately separate routine queries from demanding tasks. Anyone who habitually sends every little thing to the most capable model is wasting budget.
The merging of Chat and Work into a single interface is also not merely a cosmetic change. It reduces the effort of switching between contexts, but at the same time increases the likelihood that different types of usage will blend into one another unnoticed – and that consumption therefore becomes less transparent.
Caution when extrapolating
Reliable figures on the conversion ratio between models or on the size of the allowance are not available at this point. Anyone now fixing budgets for the coming year should therefore work with margins rather than fixed assumptions.
Only practice from December onwards will show how far a quota actually goes in typical development scenarios.
It makes sense to use the change as an opportunity for a brief review: Which tasks do AI tools actually perform in the team, what share of them does Codex account for, and which of these tasks could be solved just as well by other means?
Classification for web and software projects
For projects in which AI support has become a standard tool, tool usage thus becomes an item that must be continuously monitored – similar to hosting or licence costs.Project calculations should in future show AI usage as a variable rather than as a flat rate.And architecture decisions gain in importance: those who do not tie their processes too closely to a single product can react to changes in billing and model availability without having to rebuild the project.