OpenAI’s July 6, 2026 ChatGPT Business update makes one thing clear: workplace AI is moving from a predictable subscription experiment to a metered operating cost. Workspace Agent runs now use token-based pricing, ChatGPT for Excel/Sheets tasks do too, and PowerPoint stays free for Business customers through August 6, 2026 before it shifts to the same model.
That matters because the cost of an AI workflow is no longer just about whether it works. It is about how much context it consumes, how much it outputs, and how often your team runs it.
What changed on July 6
OpenAI’s release notes say Workspace Agent runs and Excel/Sheets tasks now use token-based credit pricing for Business and Enterprise/Edu workspaces. The rate card says pricing is measured across input tokens, cached input tokens, and output tokens, so the bill now scales with task shape instead of a fixed per-run fee.
OpenAI ChatGPT Business pricing change
| Feature | What changed | Why it matters |
|---|---|---|
| Workspace Agent runs | Token-based pricing is now in effect | Longer or more complex agent tasks cost more |
| Excel/Sheets tasks | Token-based pricing now applies | Spreadsheet-heavy workflows need tighter usage controls |
| PowerPoint | Free through August 6, 2026 | Teams have a short transition window before charges begin |
OpenAI also says Business plans include usage for Workspace Agents, Excel, and PowerPoint through the broader agentic usage pool as pricing takes effect, which makes budgeting a planning exercise rather than a guess.
Why this matters for enterprise AI budgets
Metered pricing changes behavior. Teams will start asking different questions: Which workflows are worth automating? Which tasks need long context? Which ones can be simplified so they consume fewer tokens?
That is a healthy shift, but it is also a governance problem. If an agent is doing multi-step work across files, spreadsheets, and connected apps, token usage can rise quickly. The practical risk is not just overpaying; it is scaling a workflow before anyone has measured its real cost per outcome.
For enterprise buyers, this makes agent economics look more like cloud infrastructure than software licenses. The right response is not to avoid agents. It is to budget them like compute.
What teams should do next
The first step is to audit the workflows already being tested in ChatGPT Business. Look for the ones that are repeated, high-volume, and easy to define. Those are the candidates most likely to justify a metered agent.
Then tighten the inputs. Smaller prompts, narrower files, and cleaner source data usually mean lower token usage. If a task can be broken into smaller steps, that can reduce both cost and failure risk.
Finally, decide where an agent belongs in the stack. Some jobs should stay as human-assisted chat. Others should become governed workflows with clear usage limits, review points, and ownership. The new pricing model makes that distinction more important, not less.
The bigger signal
OpenAI is not just pricing a feature. It is signaling how business AI will be sold going forward: by usage, by workflow, and by the real cost of execution. That is useful for buyers who want precision, but it also means teams need better rollout discipline from day one.
In other words, the question is no longer whether your organization can afford an AI agent. It is whether you know what that agent will cost when it starts doing real work.