Key takeaways
- The free ride ends Monday, July 6, 2026. OpenAI's workspace agents in ChatGPT — launched in research preview on April 22, 2026, and free since — start drawing down credits on that date. If you've seen "free until May 6, 2026" anywhere, that was the original deadline; OpenAI extended it on May 22, 2026, but the old date is still sitting on pages many people land on first.
- There's no flat price per run. Per OpenAI's rate card, each agent run is metered across three token streams — input, cached input, and output — with a typical end-to-end run on GPT-5.5 landing somewhere around 5 to 25 credits. Your bill depends on what your agents actually chew through, not on how many times you click "run."
- Not every run is billed — yet. Only agent runs invoked inside ChatGPT are metered from July 6, 2026. Runs triggered outside ChatGPT — an agent answering in a Slack channel, for example — stay in free preview for now. That distinction is where many teams will misjudge their exposure in either direction.
- This is the meter moment. AI agents just moved from seat pricing to utility-style metering — a first at this scale. The practical response for a one-person or small team isn't panic; it's a weekend of pruning, one honest cost estimate per workflow, and a fallback plan if the numbers surprise you.
What ChatGPT workspace agents pricing looks like after July 6, 2026
The direct answer first, because this is the question that matters: starting Monday, July 6, 2026, agent runs you invoke inside ChatGPT draw down workspace credits instead of being free. There is no fixed per-run price. Per OpenAI's rate card, each run is metered across three separate token streams — regular input tokens, cached input tokens, and output tokens — each with its own credits-per-million-tokens rate. On that rate card, a typical end-to-end agent run on GPT-5.5 works out to roughly 5 to 25 credits, depending on how much context the agent reads and how much it produces.
To make that concrete with OpenAI's own arithmetic: a run that consumes 20,000 input tokens, 80,000 cached input tokens, and 5,000 output tokens comes out to about 7.25 credits. A lightweight agent that skims a short document and posts a two-line summary sits near the bottom of the range; an agent that ingests a long thread, reasons across several tools, and drafts something substantial sits near the top — or above it.
What a credit costs you in dollars is the piece we won't pretend to pin down here. The token rates are published, but the dollar conversion, how many credits are already included in your plan, and the per-seat subscription price itself all vary by plan and billing term — OpenAI has also adjusted seat pricing over the past year, so check your workspace admin console and OpenAI's current pricing page rather than any number quoted in an article, this one included. What is structural and stable: Business, Edu, and Enterprise workspaces can purchase additional workspace credits as needed on top of the seat subscription. In other words, the seat is still flat — it's the agents that now have a meter.
If you're newer to what these agents actually do all day — triage, drafting, cross-tool chores — we've covered that ground separately in what AI agents actually automate for a small business; this brief is strictly about what those runs cost from next week.
Why you may have read "free until May 6" — the deadline that moved
Here's the trap worth naming, because a lot of people searching this topic right now are reading a stale date. The timeline, with every date in full:
- April 22, 2026 — OpenAI launched workspace agents in ChatGPT as a research preview for Business, Enterprise, and Edu plans.
- At launch, the announcement said agents would be free until May 6, 2026, with credit-based pricing starting on that date. That wording is still live on OpenAI's original announcement page — and it's echoed in the discussion threads that surface alongside it — so someone checking today can easily conclude the free window closed two months ago, or that a date already passed can't apply to them.
- May 22, 2026 — OpenAI extended the free period to July 6, 2026, announced through the ChatGPT Business and Enterprise release notes rather than a new headline post. Extensions announced in release notes travel a lot more quietly than launch announcements do.
- July 6, 2026 — credit-based billing actually begins, with the scope described below.
The practical takeaway: July 6, 2026 is the operative date, and the announcement page's "May 6" is an artifact of the original plan. If your team read the launch post in April and mentally filed agents under "already paid, apparently fine," it's worth re-checking — you've been in an extended free window, and it closes Monday.
Which runs are billed — and which stay free for now
The July 6, 2026 change has a scope line that materially changes the math for many small teams: only agent runs invoked within ChatGPT are metered. Runs invoked outside ChatGPT — the canonical example being an agent that responds in a Slack channel — remain in free preview for the time being.
Two implications, one in each direction:
You might owe less than you fear. If your heaviest agent usage is Slack-side — an agent fielding questions in a shared channel, say — that traffic isn't drawing down credits yet. Your metered exposure is only the runs that people on your team fire from inside ChatGPT itself.
You might owe more than you think. The inverse is the sneakier case. It's often the in-ChatGPT runs that are heaviest per run — long context, big documents, multi-step tool use — because that's where people do their deliberate, sit-down agent work. A small number of chunky in-ChatGPT runs can outweigh a large number of light Slack pings, and only the chunky ones are on the meter.
One caution on planning around the free side: "for now" is doing real work in that sentence. OpenAI has not committed to external invocations staying free indefinitely, so treat the Slack-side exemption as a current fact, not a permanent architecture decision. Building a workflow that's only economical because one invocation path happens to be unmetered is the kind of dependency that ages badly — a dynamic we mapped in detail in the quiet lock-in audit for AI vendors, which is exactly the playbook to reread when a vendor you're wired into starts adjusting its pricing levers.
How to estimate your own bill before Monday
You don't need a finance function to get within useful range of your July number. You need one representative workflow and twenty minutes.
Step 1 — pick your one load-bearing agent workflow. Not all of them; the one you'd actually miss. For many solo founders that's a daily digest, a lead-triage pass, or a recurring drafting job.
Step 2 — profile a single run. Run it once and be honest about what it consumes: how much context does it read (input), how much of that is repeated boilerplate the system can cache (cached input), and how long is what it produces (output)? You won't get exact token counts from eyeballing, but the 7.25-credit worked example above is a decent mental anchor: tens of thousands of input tokens plus a modest output landed at single-digit credits. A run that reads several long documents and writes pages sits meaningfully higher, toward and past the top of the 5–25 credit band.
Step 3 — multiply by honest frequency. If that agent fires once per working day, that's roughly 22 runs a month. At 5 credits a run, that's about 110 credits a month; at 25, about 550. That 5× spread is the point — flat-rate intuition ("it's included, run it whenever") is exactly the habit this pricing model punishes, and frequency is the variable you control most directly.
Step 4 — do the weekend prune. Before the meter starts:
- Inventory your agents. Preview-era experiments have a way of still being scheduled. Anything nobody would miss, retire now — an agent that was harmlessly idle at zero dollars is a small recurring leak at metered rates.
- Check trigger frequency. An agent that runs hourly "because why not" is a preview-era setting. Ask what cadence the output actually needs.
- Look at your admin console's credit and spend settings. Confirm what credits your plan includes, what happens when they're exhausted, and what spend visibility or controls your workspace offers — we're deliberately not asserting which alert features exist on which plan; verify what's in your console rather than assuming.
- Decide your fallback in advance. If August's number comes in high: lighter models where the agent's configuration allows it, lower trigger frequency, and — for low-stakes jobs — falling back to a manual prompt in an ordinary chat, since the credit meter is on agent runs, not on your regular chatting on a Business seat.
The cross-border line item: pricing is the same across the US and Canada, but billing is in US dollars. If your workspace is Canadian, your effective cost moves with the exchange rate — so budget the line in CAD with room for FX drift, and remember the invoice itself won't be in the currency your books think in.
One more piece of context worth having in the back of your mind: metering is the cost dimension of a bigger 2026 pattern in which frontier AI access itself became conditional — we covered the availability side, including the government-gated GPT-5.6 rollout, in whether Canadian businesses can use GPT-5.6 yet, and we won't re-tell that story here.
When your tools grow meters, keep one layer flat
There's a familiar feeling to July 6, 2026 if you've ever moved from an all-inclusive rent to metered utilities: nothing about the product changed, but suddenly every use has a marginal cost, and you find yourself standing in front of the meter doing arithmetic. That's the real shift here — AI agents are becoming a utility you meter, not a seat you license, and this is the first vendor of this size to flip the switch. It likely won't be the last; once one major vendor demonstrates that metered agents are commercially viable, others tend to follow.
Which makes this a good week to notice the opposite kind of cost in your business: the flat ones. Your business registration, your business address, your bank account — the identity layer of your company — doesn't meter. It costs the same whether you have a slow month or your best one, and it's the layer that makes the rest legible to banks, registries, and customers regardless of what your software stack charges this quarter. As your variable AI costs start fluctuating with usage and exchange rates, there's real value in the parts of your setup that stay predictable — if that foundation isn't settled yet, see how a North American business address works and lock in a line item that never surprises you in August.
And when you're weighing which metered tools earn a place in your stack at all, that's the standing question behind the AI tools shelf we maintain for founders — tools worth paying per-use for, versus tools that were only ever worth it free.
FAQ
Is ChatGPT Workspace paid? The workspace plans themselves have always been paid — ChatGPT Business is priced per user per month (check OpenAI's current pricing page for the going rate, which has shifted over the past year), with Enterprise and Edu priced separately. What changes on July 6, 2026 is the agents inside those workspaces: workspace agents had been free since their April 22, 2026 research-preview launch, and from July 6, 2026 agent runs invoked inside ChatGPT draw down workspace credits on top of the seat price. Ordinary chat usage on your seat isn't what's being metered — agent runs are.
Are Slack-invoked agent runs charged after July 6, 2026? Not yet. Only agent runs invoked within ChatGPT are metered from July 6, 2026; runs invoked outside ChatGPT — such as an agent responding in a Slack channel — remain in free preview for now. Treat that as a current carve-out rather than a permanent one: OpenAI hasn't committed to external invocations staying free indefinitely, so avoid building workflows whose economics only work because one trigger path is temporarily unmetered.
How many credits does a ChatGPT workspace agent run use? There's no fixed per-run price. Per OpenAI's rate card, each run is metered across input tokens, cached input tokens, and output tokens at separate credits-per-million rates, and a typical end-to-end run on GPT-5.5 lands around 5 to 25 credits. OpenAI's own worked example — 20,000 input, 80,000 cached input, and 5,000 output tokens — comes to about 7.25 credits. Heavy runs that read long documents and produce long outputs can land above the typical band.
How much is a workspace credit in dollars? The token-level rates are published in OpenAI's rate card, but a single public dollars-per-credit figure isn't something we can responsibly quote — the dollar conversion and the credits included with your plan depend on your workspace setup. Check your workspace admin console and OpenAI's rate card documentation for your actual numbers, and note that billing is in US dollars, so Canadian workspaces carry exchange-rate exposure on top of the posted rates.
Bottom line
ChatGPT workspace agents pricing flips from free preview to credit-based metering on Monday, July 6, 2026 — the "free until May 6, 2026" line still visible on the original announcement was superseded by a quiet May 22, 2026 extension. From Monday, agent runs invoked inside ChatGPT are metered across three token streams, with typical runs around 5–25 credits per OpenAI's rate card; runs triggered outside ChatGPT, like Slack-side responses, stay free for now.
The move that fits a small team: inventory your agents this weekend, profile your one load-bearing workflow against the 7.25-credit worked example, multiply by honest frequency, and check your admin console for included credits and spend controls before the meter starts. Budget in your own currency but expect a USD invoice if you're Canadian. And as more of your stack shifts from flat seats to metered usage, keep the identity layer of your business — registration, address, bank — on the flat-rate side of the ledger, where a surprise bill can't reach it.
This brief is general information for founders, not financial or legal advice. Pricing, rate cards, and free-preview scopes change quickly — confirm current terms in your workspace admin console and OpenAI's official documentation before budgeting around any figure here.
