OpenAI launched workspace agents on April 22, 2026 as a research preview for ChatGPT Business, Enterprise, Edu, and Teachers. The headline pricing detail is simple but incomplete: workspace agents are free until May 6, 2026, and then OpenAI says they will move to credit-based pricing. That means teams evaluating the feature right now need to budget in two layers: the ChatGPT workspace itself, and the agent runtime that will soon sit on top of it.
The most useful way to think about workspace agents is not as a normal seat add-on. They are closer to a managed automation layer inside ChatGPT. So the real budgeting question is not just “How much is a seat?” It is “How often will these agents run, how much work will they do, and when do we want humans in the loop?”
The short answer
Here is what OpenAI has made public so far:
- Workspace agents are in research preview for ChatGPT Business, Enterprise, Edu, and Teachers.
- They are free until May 6, 2026.
- After May 6, 2026, pricing becomes credit-based.
- ChatGPT Business is priced at $20 per user per month, billed annually, with a 2-user minimum.
- Enterprise pricing is custom.
- Business and Enterprise can purchase credits for more access.
What OpenAI has not clearly published yet on its main pricing pages is a public rate card for workspace-agent credits. That missing number is why many teams still feel uncertain even though the product is already live.
What you are really paying for
For most buyers, workspace agents will create three separate cost buckets.
1. Seat access
If your organization is not already on ChatGPT Business or Enterprise, the first cost is getting into the workspace product at all. For many smaller teams, that starts with ChatGPT Business. Larger rollouts will usually land in Enterprise, where pricing is negotiated.
2. Agent runtime
This is the new part. Once free preview ends, agent activity will consume credits. In practice, that likely means cost will scale with how often agents run, how much model work they do, how much context they pull in, and how many connected actions they take.
3. Operational overhead
Even if the direct price looks manageable, real cost also includes admin oversight, approvals, monitoring, and workflow design. Teams that skip this part often misread their budget because they focus on model spend and ignore the work required to keep automations safe and reliable.
What Budgeting Looks Like by Team Stage
Pilot phase
In an early pilot, the main cost is usually seat access plus a limited amount of future credit usage. This is the right phase to test a few narrow workflows: account research, report prep, internal ticket triage, or meeting follow-ups. The goal is not maximum automation. The goal is learning which workflows create repeatable value.
Department rollout
Once a team starts sharing agents across sales, support, finance, or operations, costs become less about experimentation and more about recurring runs. Scheduled workflows, repeated tool calls, and larger user adoption can make spend climb faster than leaders expect if they only budget for seats.
Heavy automation
If your plan is to let agents run frequent background work across tools, then the critical variable is not the sticker price of ChatGPT Business. It is total execution volume. Teams at this stage should set monthly usage caps, require approvals for sensitive actions, and separate high-frequency tasks from high-value tasks before broad rollout.
The biggest unknown: credit pricing after May 6, 2026
The main challenge for buyers is that OpenAI has announced the billing model change before publishing a clear public explanation of how many credits typical agent workflows consume. That means teams should avoid pretending they have a precise budget when they do not.
A better approach is to build a budgeting framework around uncertainty:
- Separate fixed and variable costs. Seats are the fixed layer. Credits are the variable layer.
- Count runs, not just users. A team with ten users and two daily automations may be cheaper than a team with three users and dozens of scheduled agent runs.
- Track workflow shape. Agents that summarize a document are very different from agents that search across tools, reason over multiple sources, and take actions.
- Put approval gates in the expensive paths. Not every workflow should run autonomously.
- Start with a controlled pilot cap. Treat the first month after credit billing begins as a pricing discovery period.
How workspace agents compare with API-based agent costs
Many teams are also comparing workspace agents with building directly on OpenAI’s API stack. That is a useful comparison because the API pricing is much clearer than the current workspace-agent credit model.
On OpenAI’s API pricing page, GPT-5.5 is listed at $5 per 1M input tokens and $30 per 1M output tokens. GPT-5.4 is listed at $2.50 per 1M input tokens and $15 per 1M output tokens. GPT-5.4 mini is listed at $0.75 per 1M input tokens and $4.50 per 1M output tokens. Web search is listed at $10 per 1,000 calls, and OpenAI containers are priced separately as well.
That does not mean workspace agents will cost the same as direct API usage. It does mean API-based builds give teams a more transparent cost model today. Workspace agents, by contrast, trade some pricing clarity for faster deployment inside ChatGPT and a more managed admin experience.
When workspace agents are probably the right choice
- You want employees to use agents inside ChatGPT rather than inside a custom internal app.
- You need quick deployment without building your own orchestration and UI layer.
- You care more about internal workflow acceleration than shipping a product to external users.
- You want shared agents, admin controls, and a familiar workspace experience.
When a custom agent stack may make more sense
- You need predictable usage-based cost modeling from day one.
- You want fine-grained control over models, prompts, routing, and tool policy.
- You are building customer-facing automation, not just internal productivity.
- You expect very high execution volume and want tighter control over unit economics.
What teams should do next
If you are evaluating workspace agents right now, the smartest move is to use the free preview window as a measurement period, not just a feature test. Pick two or three workflows, record how often they run, estimate how much supervision they need, and decide whether the value comes from speed, consistency, or headcount leverage. Then set a clear monthly credit budget before May 6, 2026 arrives.
OpenAI has already answered the product question: workspace agents are real, and they are clearly becoming part of the ChatGPT business stack. The open question is cost control. Until the public credit rate card is clearer, the winning teams will be the ones that budget by workflow behavior rather than by seat count alone.