TL;DR (July 2026): On July 6, 2026 the free preview of ChatGPT Workspace Agents ended. Every agent run inside a Business, Enterprise, or Edu workspace now draws down workspace credits on top of the per-seat fee you already pay. The published GPT-5.5 rates: 125 credits per million input tokens, 12.50 per million cached input tokens, 750 per million output tokens - a typical end-to-end run lands around 5 to 25 credits. What OpenAI did not publish is a consumption table for real workflows, so every team that automated prospecting, reporting, or ticket routing during the free spring now has a variable operating cost it cannot budget from the vendor's own docs. The seat is the floor; the meter is the bill. If you want a forecast, you have to build it from your own per-run usage data - nobody else has it.
The July 6 cutover is the clearest single example of the pricing shift that has been rolling through AI software all year: the subscription seat stops being the price and becomes the entry ticket, with a metered credit bill stacked on top. GitHub did it to Copilot on June 1. Microsoft finished the same move across Copilot Studio and Azure AI Foundry on July 1. OpenAI's version matters because Workspace Agents were free for the whole spring - which means thousands of teams built real workflows on a marginal cost of zero, and woke up on a Monday to find every one of those workflows metered.
What actually changed on July 6
- Before: agent runs in ChatGPT Business, Enterprise, and Edu workspaces were a free preview. The only AI line item was seats.
- After: every agent run draws down workspace credits. Credits come from plan allotments and flexible-pricing top-ups, and admins get spend controls in the workspace console.
- Unchanged: the per-seat subscription. You did not stop paying for seats; you started paying for usage on top of them.
The credit rates for the default model are public. The cost of your workflows is not - and that gap is the entire budgeting problem.
The credit math, worked
| Meter (GPT-5.5 runs) | Rate | Example: one run with 20K input + 3K output |
|---|---|---|
| Input tokens | 125 credits / 1M | 20,000 tokens = 2.5 credits |
| Cached input tokens | 12.50 credits / 1M | 20,000 cached = 0.25 credits instead |
| Output tokens | 750 credits / 1M | 3,000 tokens = 2.25 credits |
So a modest run costs roughly 5 credits, and reported typical runs land between 5 and 25. Three things in that table deserve attention before you extrapolate:
- Output is 6x input. Verbose agents are expensive agents. A run that drafts long documents or explains its work at length is billed at the 750 rate for every token of it.
- The cache discount is 10x. 12.50 versus 125 per million is the difference between a workflow that re-reads the same context cheaply and one that pays full freight every run. Whether your runs hit the cache is invisible unless you track it - the same trap we documented in how prompt caching quietly breaks cost tracking.
- A "run" is not a unit of work. The same agent asked the same question twice can read different amounts of context, call different tools, and produce runs that differ several-fold in credits. That variance is exactly why coding agents moved to session and task quotas - and why a per-run average from one week is a shaky forecast for the next.
Why teams cannot budget this from the docs
The complaint circulating among buyers is not the rates - it is that the meter shipped before the consumption table. OpenAI published what a token costs in credits, but not what a workflow costs in tokens. Nobody outside your workspace knows how many tokens your sales-prospecting agent reads per account, or how much output your weekly finance-report agent produces. The vendor cannot publish your workflow's cost because the vendor does not know it. Only your own meter does.
This is the same lesson the Copilot cutover taught in June, at larger scale: when GitHub switched 4.7 million subscribers to AI Credits, the users who got hurt were the ones who discovered their consumption profile from the bill. Heavy agentic users reported multiples of their old cost, not because the rates were hidden, but because nobody had measured their own usage while it was free.
The free-preview trap, named
A free preview sets a marginal cost of zero, and teams rationally build as if that is the price. Every workflow adopted in the free window embeds an unmeasured consumption profile. Then the meter turns on, and the finance question - "what will Workspace Agents cost us in August?" - has exactly one honest answer: whatever your runs consume, and you never measured your runs.
The pattern to internalize: the time to meter a free AI feature is while it is free. Usage data collected during the preview is the only thing that turns a metered cutover from a surprise into an arithmetic problem. We made the same argument about free API tiers in the free AI APIs that actually stay free - free is not unmetered, it is a meter you are not watching yet.
What to meter now (whether or not you use UsageBox)
- Credits per run, tagged by agent. The workspace bill aggregates; your meter should attribute. A month of runs tagged by agent name gives you the consumption table OpenAI did not publish.
- Runs per workflow per week. Volume is the other half of the forecast. Credits-per-run times runs-per-week is your August number.
- Who invoked what. Agent runs triggered by individuals, schedules, and integrations look identical on the invoice. Attribution by invoker is how you find the one automation that quietly runs hourly - the same discipline as cost attribution across tools, sub-agents, and retries.
- Cache hit rate per workflow. At a 10x discount, cache behavior is a first-order cost driver, not an optimization detail.
- A spend threshold with an alert. Admin spend controls cap disasters; a meter with an alert catches drift the week it starts, not the month after.
The honest take
Seat-plus-meter is not a trick - it is the correct shape for a product whose cost to serve varies 100x between users. Flat seats subsidized heavy users; the meter ends the subsidy. But the burden it shifts onto buyers is real: a credit meter is a utility bill, and managing it is closer to managing cloud compute than managing a SaaS contract. The teams that will do this well are the ones that treat their own usage data as the primary source of truth - because as of July 6, for Workspace Agents, it is the only source of truth that exists.