July 6, 2026. The free ride for ChatGPT workspace agents ends today. OpenAI's ChatGPT rate card sets July 6, 2026 as the effective date for credit based pricing on Workspace Agent runs for Business and Enterprise plans. Every scheduled report builder, lead qualifier, and month end close agent that teams wired up during the free preview now has a per run price. For any business that spent the spring handing real jobs to agents, today is the day the meter starts, and the difference between an agent that pays for itself and one that quietly burns budget comes down to how it was built.
Workspace agents are OpenAI's successor to custom GPTs: Codex powered agents that run in the cloud rather than inside a chat session, work on schedules, keep memory across projects, and connect to tools like Slack and Salesforce. They launched April 22 in research preview with free usage, OpenAI declared them generally available on May 22 and extended the free window from the original May 6 cutoff to July 6, and Tech Times flagged the hard date back in June. We covered the launch in our June 22 workspace agents brief and the broader metered AI shift in our July 2 agency cost playbook. Today's piece is the agent level follow through: the actual math of a single run, and what to do about it before the first invoice.
The key developments
- Credit pricing for ChatGPT Workspace Agents takes effect today, July 6, 2026, per OpenAI's rate card. The free period that began with the April 22 launch, originally set to end May 6 and then extended on May 22, is over for runs invoked inside ChatGPT.
- Runs are metered on tokens, not priced per run. GPT-5.5 agent runs consume 125 credits per million input tokens, 12.50 credits per million cached input tokens, and 750 credits per million output tokens. A typical end to end GPT-5.5 run lands between 5 and 25 credits.
- OpenAI's own worked example: a run using 20,000 input tokens, 80,000 cached input tokens, and 5,000 output tokens costs about 7.25 credits (2.5 for input, 1 for cached input, 3.75 for output).
- GPT-5.4 runs cost exactly half the GPT-5.5 rates (62.50 input, 6.25 cached, 375 output per million tokens), and requests ChatGPT auto routes to a mini model consume zero credits.
- There is a carve out: the new rates apply to agent runs invoked within ChatGPT. Runs invoked outside ChatGPT, such as agents replying in connected Slack channels, remain in free preview for now, and OpenAI has not published a firm end date for that preview.
- There is no public dollar price per credit. Business workspaces buy pooled credit packs (valid 12 months) through Settings and Billing; Enterprise and Edu customers purchase a shared credit pool through their OpenAI account team, with allocation set in the order form. Business plans also include some agent usage drawn from the shared Codex agentic usage pool before flexible pricing kicks in.
How a single agent run is priced
The unit economics are simple once you see the three meters. Every run reads fresh input tokens at full price, reads cached input tokens at one tenth of that price, and writes output tokens at the most expensive rate of the three. Task complexity, input size, cache hits, and output length decide where a run lands in the 5 to 25 credit range. That fivefold spread is the whole game: the same business workflow can sit at either end depending on how it is architected.
The cached input rate is the design lever that matters most. Cached input on GPT-5.5 is billed at 12.50 credits per million tokens versus 125 for fresh input, because the model reuses context it has already processed: instructions, knowledge files, and conversation history that have not changed between runs. An agent with stable instructions and a stable knowledge base rides the cheap lane on every scheduled run. An agent that re ingests large fresh documents on every run pays full price, every time. If you build or buy agents, this is now a line item question, and it is the kind of decision an experienced AI engineer makes at design time, not after the first surprising bill.
The day one playbook
First, inventory every published agent by invocation surface. Agents that answer in Slack are free for now; agents that run on schedules inside ChatGPT start consuming credits today. The admin console shows each agent's total runs, unique users, and usage over time, and the Compliance API exposes every agent's configuration and run history, which is the raw material for a cost per workflow estimate.
Second, price your top agents with the formula. Take a representative run, estimate the three token streams, and apply the rates. A daily agent averaging 25 credits per run compounds to roughly 9,000 credits a year on its own, before you multiply across a fleet of shared agents.
Third, restructure the expensive ones. Move stable context (instructions, reference files) into the agent so it caches, trim what the agent re reads fresh each run, cap output length where a summary does the job, and downgrade agents that do not need frontier reasoning to GPT-5.4 at half rate. Routine work should not default to the most expensive path.
Fourth, set the guardrails before the first bill. Business admins can configure usage alerts and automatic recharge, or leave recharge off so advanced features pause when the pool empties. Enterprise owners can set hard overage limits (a limit of zero blocks overages entirely) and role based spend rules by group.
Fifth, decide with ROI, not sticker shock. OpenAI's launch post quoted Rippling running a sales opportunity agent that researches accounts, summarizes calls, and posts deal briefs into Slack, replacing five to six hours of weekly rep work per deal. Against loaded payroll, a few hundred credits a month for that outcome is a rounding error. The agents to kill are the orphaned experiments nobody owns, not the ones doing measurable work.
What it means for operators
Metered agents reward teams that measure and quietly penalize teams that do not. That is the same lesson the whole stack taught this month, from Copilot's July 1 billing cutover to Claude's metered programmatic usage, and it changes what good automation looks like: tightly scoped agents, stable cached context, the smallest model that clears the quality bar, and a human owner per agent who knows its cost per outcome. For most small and mid sized businesses the constraint is not the credits, it is the design skill. That is exactly the work of an AI automation partner: scope the workflow, build the agent so its unit economics hold, and instrument it so you can prove what it earns. Agencies running agent fleets for clients should treat today as a repricing event, per our July 2 playbook, and an AI automation agency that already tracks cost per workflow will have no trouble showing clients the math.
The bigger signal: agent pricing is converging on token metering across every vendor, which means the durable advantage is not access to any one platform, it is owning well designed workflows whose value per run beats their cost per run. Build for that and the meter is your friend, because it makes the ROI visible.
Frequently Asked Questions
OpenAI's rate card sets July 6, 2026 as the effective date for Workspace Agent credit pricing on Business and Enterprise plans. The free period that started with the April 22 launch, and was extended once from May 6, ends on that date for runs invoked inside ChatGPT.
Runs are metered on tokens, not per run. On GPT-5.5, input tokens cost 125 credits per million, cached input tokens 12.50 per million, and output tokens 750 per million. OpenAI's example run (20,000 input, 80,000 cached, 5,000 output tokens) costs about 7.25 credits, and a typical run lands between 5 and 25 credits.
Not yet. The July 6 rates apply to Workspace Agent runs invoked within ChatGPT. Runs invoked outside ChatGPT, such as agents responding in connected Slack channels, remain in a free preview, and OpenAI has not published a firm end date for it, so treat that carve out as temporary.
OpenAI does not publish a public dollar price per credit. Business workspaces buy pooled credit packs through Settings and Billing that are valid for 12 months, while Enterprise and Edu customers buy a shared credit pool through their OpenAI account team with terms set in the order form.
Ride the cache: keep instructions and knowledge files stable so repeat context bills at the cached rate, which is one tenth the fresh input rate. Also trim what the agent re reads each run, cap output length, and move agents that do not need frontier reasoning to GPT-5.4, which is billed at half the GPT-5.5 rates.
Only the orphans. Agents doing measurable work usually clear the bar easily; OpenAI cited a Rippling sales agent replacing five to six hours of weekly rep work per deal, which dwarfs its token cost. Inventory every agent, compute cost per outcome, keep what earns, restructure what is expensive, and retire what nobody owns.