On July 9, 202601, OpenAI shipped ChatGPT Work. The same day, it launched the GPT-5.6 family of models, which powers ChatGPT generally. Reuters summed up the pitch in one phrase: coding-tool power for white-collar workers "without the sticker shock."
Three days earlier, on July 6, OpenAI started charging for workspace agent runs.
Look at the order of operations. Workspace agents opened as a research preview on April 22, 202613. Free usage was supposed to end May 6. Then the meter turned on, and three days later the flagship consumer-facing version of the same capability began rolling out in Pro, Enterprise, and Edu, with Plus and Business following within days.
That is not a product roadmap. That is an offer being built in public.
The Seat-to-Meter Flip
Here is the framework: give the run away, wire it into the workflow, then meter the run.
Four numbers that show the seat-to-meter flip already happened.
Software has sold seats for thirty years. A seat is a great business because it is predictable. A seat is also a terrible weapon, because your competitor can always sell a cheaper seat. Microsoft sells Copilot seats. Google sells Workspace AI seats. Both are priced per human.
OpenAI is pricing per unit of work instead. According to OpenAI's rate card reported by Tech Times on June 10, 202610, a typical GPT-5.5 workspace agent run consumes 5 to 25 credits, with no fixed per-run price. That is a 5x spread between a cheap run and an expensive one, decided by how hard the task is. Runs invoked from Slack stayed free at launch, which tells you distribution still matters more than margin.
One clarification before we go further. The API price cuts are real, but they landed separately from the agent launch: on July 30, 2026 OpenAI cut GPT-5.6 Luna by 80 percent to $0.20 per million input tokens and Terra by 20 percent to $2 per million input tokens, leaving flagship Sol unchanged. Three weeks separate them from the July 9 launch, so the metered agent run, not the rate card, is the load-bearing half of the thesis.
The Golden Goose Is The Run, Not The Seat
Every pricing decision answers one question: what exactly am I selling?
Most AI companies think they are selling access. Access is a golden egg. It gets cheaper every quarter and someone will always undercut you. The golden goose is the completed unit of work, because a completed unit of work gets compared against something other than a rival vendor. It gets compared against a salary.
That reframe is the entire offer. Nobody can tell you whether $30 a seat is expensive. Everybody can tell you whether a finished five-page report is worth more than the two hours it used to eat. Price per run and you stop arguing with Copilot's price list. You start arguing with payroll. That is a much better fight to pick.
The 75-day free window was risk reversal, not generosity. In offer terms, you remove the price so the only remaining objection is "does it work." Once a team has built three workflows on top of agent runs, connected them to their files, and shared them across a department, the objection is gone and so is the alternative. Switching cost is not built by contracts. It is built by habit.
Then the pricing model does the compounding. A seat license caps your revenue at headcount. A credit meter has no cap. If enterprise usage doubles, revenue roughly doubles, with no new logos, no new salespeople, no renewal negotiation. OpenAI's Chief Revenue Officer Denise Dresser wrote on April 8, 202616 that enterprise had already crossed 40% of revenue and was tracking toward parity with consumer by the end of the year. She also disclosed Codex at 3 million weekly active users and APIs processing more than 15 billion18 tokens per minute.
Now the damaging admission, because this is where the offer gets fragile. Consumption pricing cuts in both directions. An agent stuck in a loop bills credits with the same enthusiasm as an agent doing brilliant work. Security researchers at Zenity Labs demonstrated a vulnerability they called AgentForger, where a single malicious link could schedule a rogue workspace agent inside a victim's account; OpenAI patched it on June 8, 202616. A finance team that gets one surprise invoice, or one security incident, will cap spend faster than any competitor can undercut it.
I think that is the real risk to the land-grab, and it is not competitive. It is a CFO with a spreadsheet and a bad month.
Three signals inside the same shift
The unit of sale moved from the human to the task.
Microsoft and Google price Copilot and Workspace AI per human. OpenAI prices per completed run at 5 to 2510 credits, with no fixed per-run price. That stops the comparison against a rival price list and starts a comparison against payroll.
The free window was risk reversal, not generosity.
Workspace agents opened as a research preview on April 22, 202613 with free usage supposed to end May 6. The extension removed price as an objection until teams had wired agents into files and workflows. Switching cost got built by habit, not by contract.
The meter cuts both ways.
An agent stuck in a loop bills credits as enthusiastically as one doing brilliant work. Zenity Labs demonstrated AgentForger, where a single malicious link could schedule a rogue workspace agent inside a victim account. One surprise invoice caps spend faster than any competitor can undercut.
Reprice one thing you sell from access to outcome.
- Name your golden goose. Write down what you actually charge for today, then write down the completed unit of work your buyer cares about. If those two sentences do not match, you are selling access and someone will undercut you next quarter.
- Build a credit spread, not a flat fee. Copy the 5 to 2510 band: cheap tasks bill light, hard tasks bill heavy, and the price is decided by difficulty rather than by a price list. Keep one distribution channel free the way OpenAI kept Slack-invoked runs free at launch.
- Pre-empt the CFO spreadsheet. Ship spend caps, per-run cost visibility, and a loop detector before you turn the meter on. Then run the AgentForger question on your own product: what happens if one malicious link schedules a run inside a customer account?
The competitor is not Copilot. It is a CFO with a spreadsheet
A seat license caps revenue at headcount, and a credit meter has no cap, so if enterprise usage doubles the revenue roughly doubles with no new logos and no renewal fight. That is why the July 6 metering switch matters more than the July 9 launch headline, and why the July 30 API price cuts on Luna and Terra, real but three weeks removed from the launch, remain the weaker half of the thesis. Dresser's disclosure that enterprise already crossed 40%16 of revenue, alongside 3 million weekly Codex users and 15 billion18 tokens per minute, shows the volume side is working. The fragile side is trust in the invoice. Metered work only compounds if the buyer never has a bad month.