K Koda Intelligence
exploreDeep Dive
DEEP DIVE BRIEFING № 126 · 31 July 2026
Live Intelligence Fact-checked

OpenAI is no longer selling seats. It is selling finished work

On July 9, 2026 OpenAI shipped ChatGPT Work alongside the GPT-5.6 family. Three days earlier, on July 6, it quietly started charging for workspace agent runs that had been free since the April 22 research preview. A typical GPT-5.5 agent run now burns 5 to 25 credits, priced by task difficulty rather than by headcount. That sequence is not a roadmap. It is an offer built in public.

7 MIN READ · BY THE KODA EDITORIAL TEAM · STRATEGY · ENTERPRISE PRICING
JUL 9CHATGPT WORK· REUTERS
JUL 6METER ON↑ OPENAI
APR 22AGENT PREVIEW· RESEARCH PREVIEW
keyboard_arrow_down
graphic_eq
LISTEN · AUDIO BRIEFINGThe conversation · ~2 min
smart_display
WATCH · VISUAL NARRATIVEAnimated breakdown · ~2 min
play_arrowPLAY · YOUTUBE
CHATGPT WORKJUL 9· REUTERS METER ONJUL 6↑ OPENAI AGENT PREVIEWAPR 22· RESEARCH PREVIEW CREDITS PER RUN5-25↑ TECH TIMES FREE WINDOW75 DAYS↓ RISK REVERSAL ENTERPRISE MIX40%↑ OF REVENUE CODEX WAU3M↑ DRESSER API THROUGHPUT15B↑ TOKENS / MINUTE

On July 9, 2026, 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, 2026. 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.

PRICING LEDGER · JULY 2026REUTERS · TECH TIMES · OPENAI · ZENITY LABS

Four numbers that show the seat-to-meter flip already happened.

Enterprise share of revenue Denise Dresser · April 8, 2026
40%
Credit spread per agent run Tech Times rate card · June 10, 2026
5-25
Codex weekly active users OpenAI disclosure · April 2026
3M
API tokens per minute OpenAI disclosure · April 2026
15B

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, 2026, 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 honest caveat before we go further. The story circulating is that OpenAI paired the agent launch with sharp API price cuts. I could not verify a published before-and-after API rate table for July 2026 in the primary sources. What is documented is the metering change, the extended free period, and the tier structure of GPT-5.6 across plans. So treat "API price cuts" as the weaker half of the thesis and the metered agent run as the load-bearing half.

The Golden Goose Is The Run, Not The Seat

Every pricing decision answers one question: what exactly am I selling?

Coding-tool power for white-collar workers without the sticker shock.· REUTERS · JULY 9, 2026

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, 2026 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 billion 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. 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

METER OVER SEAT
5-25

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 25 credits, with no fixed per-run price. That stops the comparison against a rival price list and starts a comparison against payroll.

FREE THEN BILLED
75 DAYS

The free window was risk reversal, not generosity.

Workspace agents opened as a research preview on April 22, 2026 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.

CONSUMPTION RISK
JUL 6

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.

DOJO · BUILD THIS WEEKEND

Reprice one thing you sell from access to outcome.

  1. 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.
  2. Build a credit spread, not a flat fee. Copy the 5 to 25 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.
  3. 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?
Train the full skill in The Dojoarrow_forward
THE BOTTOM LINE

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 reported API price cuts remain the weaker half of the thesis with no published before-and-after rate table in the primary sources. Dresser's disclosure that enterprise already crossed 40% of revenue, alongside 3 million weekly Codex users and 15 billion 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.

EDITORIAL RECEIPTKODA-20260731-910B197BB0C2
As of31 July 2026MethodClaim extraction, dated-evidence review, and temporal consistency gate.CorrectionsContact the Koda desk
Filed underStrategyDeep Dive31 July 2026
Browse the Deep Dive archivearrow_forward

Want this every morning?

AI analysis, world news, markets, and tools. One briefing, delivered free.

One email per day. No spam. Unsubscribe anytime.