Workspace Agents OpenAI on credit from May 6, 2026: what’s changing for teams

Since May 6, 2026, Workspace Agents OpenAI are no longer free.

The open test phase that began on April 21 closed at midnight Pacific time.

Every action triggered by an agent now consumes credits deducted from the ChatGPT Business, Enterprise, Edu, and Teachers plans.

The critical budget detail: as of May 10, 2026, OpenAI has not published any public rate card.

Four days after the switch, teams are paying blindly.

This audit reconstructs the price range from comparable public grids, estimates a comparison for 1,000 monthly tasks against five alternatives, highlights the tax black hole for French companies, and lists five tactics to apply before the next invoice.

What changes concretely on May 6, 2026

The official timeline consists of three dates.

OpenAI announced the Workspace Agents on April 22, 2026, opened a free research preview immediately, and confirmed the switch to credit billing on May 6 at midnight Pacific time (3 AM Eastern, 9 AM French time).

Four days later, on May 10, 2026, the official rate card is still missing on openai.com and in the help center.

OpenAI started billing without publishing the minute price, the budgetary equivalent of a taxi with a hidden meter: every action of the Workspace Agents OpenAI runs, but no one sees the rate until the end-of-month invoice.

Official timeline as of May 10, 2026

On May 5, 2026, OpenAI updated the product page to confirm the switch without detailing the pricing.

On May 6 at 00:00 Pacific, the meter started for eligible accounts.

As of May 10, several independent tech sources (FindSkill, OpenTools, FutureAIStack) all confirm the same observation: the per-credit grid is neither in the help center, nor in the billing terms, nor in response to support tickets.

Any estimate circulating today is an extrapolation, not an official rate.

Eligible plans and seat pricing

The feature remains reserved for pro plans.

ChatGPT Business is billed at $25 per user annually or $30 monthly.

Enterprise is around $60 per user based on public estimates.

Edu and Teachers maintain their specific pricing with uncertainty for K-12 American teachers who had free access until 2027: OpenAI has not confirmed if agent credits are included.

The Plus, Pro, and Free plans never had access to Workspace Agents and are not affected by the switch.

How much it will cost: the range OpenAI doesn’t publish

Without an official rate card, two public references serve as a guide.

The Microsoft Copilot Studio grid charges $200 for 25,000 credits in a monthly capacity pack, which is $0.008/credit, or $0.01/credit in pay-as-you-go.

The Notion Custom Agents grid charges $10 for 1,000 credits, exactly $0.01/credit.

The floor-to-ceiling range set by these two players frames OpenAI’s probable zone: between $0.008 and $0.01 per credit.

This extrapolation remains an estimate, to be recalibrated once the official grid is released.

Extrapolation from copilot studio and notion

The Copilot Studio grid is the most informative: pack of 25,000 credits at $200/month, overage at $0.01/credit in PAYG, autonomous triggers billed at 25 credits each.

To delve into this model, see the credit pricing of Microsoft Copilot Studio, which details the mechanism and its pitfalls.

Notion publishes a simpler grid: 1,000 credits for $10, with no rollover, shared at the workspace level.

The feature became paid on May 4, 2026, two days before OpenAI.

Estimation by use case

An independent estimate published by FindSkill positions an agent at 100 invocations per day on standard GPT-5.4 at about $300/month.

The calculation: 3,000 monthly invocations, about 30 credits per run for a multi-step agent, at $0.01/credit, gives $900/month on Notion or $720/month on Copilot Studio in PAYG.

The projection at $300 assumes a lighter agent or use of the capacity pack at a reduced rate.

The clearest leverage comes from the choice of model.

Notion documented in April 2026 a 35 to 50% reduction on Custom Agents switching to GPT-5.4 Mini, Nano, Haiku 4.5, or MiniMax M2.5.

For repetitive workflows, credit consumption can drop by a factor of 10.

End-to-end cost comparison for 1,000 tasks per month

To make the subject concrete, here is the projection for the same volume: 1,000 agent tasks per month, the profile of a 12-person web agency or an SME automating its monitoring, support, and draft generation.

The figures are projections as of May 10, 2026, to be recalibrated upon the release of the OpenAI rate card.

An honest comparison doesn’t rank tools by unit price but by total included cost: an agent at $0.03 per run requiring 80 hours of dev at the start is more expensive than an agent at $0.15 activated in 30 minutes for most teams.

Workspace agents versus four alternatives

For 1,000 monthly tasks, the projected orders of magnitude fit within this range:

• Workspace Agents OpenAI: between $240 (low estimate, Mini model) and $2,400 (high estimate, standard GPT-5.4 and 60 credits per run), plus the cost of Business seats at $25 × seats.
• Claude Cowork: Claude Team at $25/seat × 12 = $300 for seats, plus API consumption depending on the model. To explore the mechanics, see Anthropic’s managed agents (Claude Cowork).
• Microsoft Copilot Studio: $200 for 25,000 credits, providing ample coverage for 1,000 runs around 25 credits each.
• Notion Custom Agents: $10 for 1,000 credits, equating to 100 to 333 runs depending on the agent, around $30-100 for 1,000 light runs, up to $200-300 for heavy runs.
• Mistral Le Chat: fixed-cost Team packages, plus Mistral Large or Medium API consumption, the most budget-predictable profile with a more limited number of integrations.

When internal build becomes cheaper

The direct OpenAI API disrupts the projection.

10 million tokens per month on GPT-5.4-mini cost about $50 via the direct API, compared to the projected $16,000 to $20,000 through Workspace Agents for the same volume of tasks.

The switch becomes profitable from 5,000 to 7,500 tasks per month.

At this volume, an in-house agent via LangGraph or CrewAI plus an API running on Claude, Mistral, or GPT-5.4-mini drops to around $500 to $1,000/month.

The hidden cost: 500 to 2,000 hours of dev to reach the reliability of a managed product.

The threshold shifts depending on the team’s profile and workflow complexity.

To compare foundational subscriptions, also see the Claude and ChatGPT 2026 pricing comparison.

Tax black hole for French companies

OpenAI does not communicate on the legal entity billing ChatGPT Business subscriptions from France.

Three entities exist in OpenAI’s public documentation: OpenAI Ireland Ltd (Dublin, D01 YC43), OpenAI France SAS (95 Rue La Boétie, Paris 75008), and OAI Ireland Limited for countries without a local entity.

The invoice submission guidelines page specifies that the invoice must be addressed to the entity referenced on the purchase order, without automation.

The concrete risk for an SME or freelancer: receiving an invoice issued by OpenAI Ireland Ltd, without a French SIRET, with a reverse-charge mention or no VAT mention, and having to manually reconstruct the accounting document to justify the deduction.

The legal mechanism is clear: for a B2B service provided by an EU supplier not established in France, the French client applies the VAT self-assessment (article 283-2 of the CGI), provided they have a valid intra-community VAT number verifiable on VIES.

The invoice must show 0% VAT and the mention “Self-assessment” or “Reverse charge — Article 196, Directive 2006/112/EC”.

Self-employed individuals exempt from VAT are most exposed: without an intra-community VAT number, OpenAI may either refuse the zero-rated invoice or charge Irish VAT (23%) which is not recoverable in France.

Checking the issuing entity on the first invoice of the May cycle and keeping a VIES screenshot are two non-negotiable reflexes.

5 tactics to contain the bill

With a hidden meter, a defensive approach is paramount.

The five tactics below can be applied this week, without waiting for the official rate card.

They are based on documented feedback from Notion, Copilot Studio, and the HN community between April 22 and May 9, 2026.

Model selection: Mini or Nano by default

The simple rule: Mini or Nano by default, GPT-5.4 standard reserved for steps requiring reasoning (strategic synthesis, final client deliverable generation).

Notion documented up to 10 times fewer credits per run by switching to lighter models.

On an email triage agent with 600 monthly runs, the difference adds up quickly: $180 in GPT-5.4 standard versus $18 in Nano.

The reflex to industrialize: audit each step of each agent, identify those that don’t require the top model, switch without delay.

Batching and caching: anticipate a short TTL

Anthropic reduced the TTL of its prompt cache to 5 minutes in early 2026.

OpenAI is likely to follow.

Batching groups multiple requests under the same context to amortize the cost of embedding and reasoning.

Caching stores the response to identical prompts to avoid a new charge.

On a workflow calling the same client doc synthesis 30 times a day, caching reduces consumption by 20 times on repeated requests.

Structuring agent prompts to leverage these two levers now avoids technical debt when OpenAI adjusts its cache.

Daily monitoring and alert at 70%

A daily updated agent tracking dashboard is the minimum safety net.

Automatic alert at 70% of expected consumption: an agent going off track is seen in a few hours, not at the end-of-month invoice.

On Notion, cases reported on Reddit describe workspaces consuming 150,000 credits in a month ($1,500) before anyone noticed.

The granularity of tracking is as important as the threshold: trace by agent and by model to identify the source of the overrun in 5 minutes.

Conclusion

May 6, 2026 marks the end of a free period that many teams took for granted.

Without a public rate card, the 2026 budget arbitration on AI agents is blind for a few weeks, until the first invoices arrive and the community triangulates the real figures.

The Workspace Agents OpenAI are not a trap, but their opening without a pricing grid requires an auditor’s stance: measure, compare, switch to Mini or Nano by default, and reserve the possibility of an internal build beyond 5,000 monthly tasks.

Preparing the 2026 AI budget without surprises starts by marking your own thresholds, cross-referencing the first invoices against Copilot Studio and Notion estimates, and arbitrating model by model before the June cycle invoice.

FAQ

How much will a Workspace Agents OpenAI agent cost per month after May 6, 2026?

The extrapolated estimate from public grids Microsoft Copilot Studio ($0.008/credit) and Notion ($0.01/credit) places an agent at 100 invocations per day around $300/month in standard GPT-5.4.

The range drops to $30-60/month when switching to Mini or Nano models.

Is the OpenAI credit price published as of May 10, 2026?

No, OpenAI has not published any official rate card four days after the switch.

Any estimate circulating today is an extrapolation from Copilot Studio and Notion, not a confirmed rate.

Do ChatGPT Business and Enterprise plans include free agent credits each month?

OpenAI has not confirmed any included allocation.

The model followed at this stage resembles pure pay-as-you-go on the invoice, without an explicit free pool, to be verified on the first invoice of the May cycle.

My agent that ran for free in April, what will I pay in May?

Three light agents running three times a week project between €80 and €250 in GPT-5.4 depending on complexity, compared to €0 in April.

Switching to Mini or Nano divides this projection by 5 to 10.

Which alternative is the cheapest for 1,000 tasks/month?

Microsoft Copilot Studio at $200 for 25,000 credits comfortably covers 1,000 heavy runs.

Notion Custom Agents holds around $30-100 for light runs; Mistral remains competitive on the fixed package plus API.

From what volume does an open-source in-house agent become cheaper?

The quantified threshold is around 5,000 to 7,500 tasks per month.

Beyond that, a LangGraph or CrewAI agent plus the GPT-5.4-mini, Claude, or Mistral API drops to $500-1,000/month compared to the projected $15,000 to $20,000 on Workspace Agents.

Will I receive a VAT-inclusive invoice with my SIRET or an untaxed invoice from OpenAI Ireland?

OpenAI does not specify the default issuing entity.

French companies identified for VAT generally receive a zero-rated invoice from OpenAI Ireland Ltd, to be self-assessed on the CA3 with the mention “Self-assessment”; self-employed individuals exempt from VAT are exposed to non-recoverable Irish VAT.

How to monitor daily agent credit consumption?

The minimum is a tracking dashboard by agent and model, with an automatic alert at 70% of expected consumption.

Trace each run to identify peaks before they impact the monthly invoice.

Will Anthropic’s 5-minute TTL caching affect OpenAI too?

The hypothesis is strong: LLM cache is expensive on the infrastructure side, and OpenAI has already adjusted its runtime parameters several times without notice.

Structuring agent prompts to exploit short-term cache now avoids technical debt at the next adjustment.

Can I quickly migrate from Workspace Agents OpenAI to Claude Cowork or Microsoft Copilot Studio?

Not without cost.

Each alternative imposes its own integration: Slack connectors, prompt formats, memory mechanics, requiring 20 to 80 hours per agent for a clean migration, plus parallel testing for two to four weeks.