The seat is dead, long live the credit: OpenAI quietly retires Custom GPTs

OpenAI launched workspace agents in ChatGPT this morning. The announcement¹ is framed as a product upgrade, Codex-powered, persistent, Slack and Salesforce integrations, memory across sessions, the works, but the thing I want to look at is not the capability story. It is the pricing page, and the sentence in the FAQ about Custom GPTs, and the way these two things together complete a migration OpenAI has been running for about eighteen months.

The migration is off seats and onto credits. Custom GPTs, the no-code builder that shipped in November 2023 as the flagship of the ChatGPT Team and Enterprise value prop, are being parked. They continue to work. You can still use them. But "teams can migrate them"² to workspace agents, which run on credit-based pricing after a two-week free window closing May 6. The documentation is polite about this. The pricing is not: Custom GPTs were included in the per-seat subscription. Workspace agents are not.

Custom GPTs, the no-code builder that shipped in November 2023 as the flagship of the ChatGPT Team and Enterprise value prop, are being parked.

This is the SaaS apocalypse frame firing almost too cleanly to be interesting, except that the mechanism is specific enough to be worth walking through.

What the frame predicts. Per-seat SaaS pricing assumes a human user doing work at roughly human pace. When the unit of work shifts from human-hours to agent-invocations, when the thing consuming the software is an agent that can run a hundred workflows a night, seat pricing breaks in both directions. It under-prices heavy agentic users (who get unbounded inference on a fixed subscription) and over-prices light users (who pay for a seat they barely invoke). The frame predicts that vendors with exposure to inference costs will move, first quietly and then not quietly, from seats to consumption.

What OpenAI actually did. ChatGPT Team remains $30 per seat per month.³ The seat still exists. What has changed is what the seat entitles you to. Under the old regime, a seat bought access to Custom GPTs, bounded agents, essentially templated prompts with retrieval, at no marginal cost. Under the new regime, the seat buys access to the workspace agents interface, and the agents themselves bill in credits. OpenAI has not retired the seat. It has hollowed it out. The seat is now a license to spend money, rather than a license to compute.

This is the move. It is the same move Microsoft made with Copilot Studio's "messages" billing in 2024,⁴ and the same move Salesforce made with Agentforce's per-conversation pricing. It is the move every vendor whose cost of goods sold is dominated by inference is going to have to make, because the alternative is that your heaviest-value customers are also your biggest gross-margin holes.

The Custom GPT question. Here is where it gets slightly amusing. Custom GPTs were, by OpenAI's own telling in late 2023, the foundation of the enterprise strategy. Sam Altman on stage at DevDay: "GPTs are a new way for anyone to create a tailored version of ChatGPT." The GPT Store was announced as a revenue-share marketplace. There was, briefly, a narrative in which Custom GPTs were going to be the App Store of AI.

The GPT Store is not mentioned in today's announcement. The revenue share, which was rolled out to a small set of US builders in Q1 2024, is not mentioned. The FAQ says Custom GPTs "will continue to be available" and that "we recommend exploring workspace agents for more capable, persistent workflows."⁵ This is the softest sunset I have read in a while. The feature isn't being killed. It is being made uninteresting.

I notice this partly because sunsetting-by-neglect is the cleanest possible tell on unit economics. If Custom GPTs were margin-positive at the seat price, OpenAI would be building on them, not routing around them. They are being routed around. Draw your own conclusions about what a bounded GPT costs to serve versus what a $30 seat brings in.

The credit mechanic. The pricing page gives credit allotments per plan tier and a per-credit overage rate, but does not disclose the credit-to-inference mapping.⁶ This is standard for the category now, Anthropic's Claude for Work does the same, as does Google's Gemini Enterprise, and it is worth naming what it is. The vendor retains the right to re-price inference against the credit without re-pricing the subscription. If a Codex run that costs 10 credits today costs 14 credits in six months because GPU supply tightened or the model got heavier, the customer sees a usage increase, not a price increase. The SKU is stable. The economics underneath it are not.

For a strategist evaluating this, the question is not whether workspace agents are good. They probably are; Codex in agent form with persistent memory and enterprise integrations is a genuinely useful product. The question is what your inference bill looks like in twelve months when the agent you deployed to triage support tickets is running continuously and the credit mapping has drifted twice.

What this is a case of. Three things, stacked:

First, it is a case of a vendor with exposure to inference costs completing the transition off seat-based pricing for the agentic tier while preserving the seat SKU as a billing anchor. Expect Anthropic, Google, and the Microsoft Copilot stack to converge on this shape if they haven't already.

Second, it is a case of a first-generation AI product feature (Custom GPTs) being deprecated not because it failed but because its pricing model didn't survive contact with actual usage. The GPT Store, insofar as it existed, was a per-seat marketplace. The workspace agents equivalent, if it materialises, will be a credit-share marketplace. That is a different business.

Third, and this is the one I would watch, it is a case of the seat becoming a line item rather than a pricing unit. Once the seat is just the access ticket and all the value flows through consumption, the strategic question for buyers becomes whether to negotiate seats and credits together or separately. Procurement teams have not caught up to this yet. They will.

What to watch. The May 6 cliff, when free usage ends, specifically, whether OpenAI publishes any utilisation data or whether pricing holds through Q3. Whether the GPT Store gets a public update or quietly disappears from the product surface. Whether Microsoft 365 Copilot, which has been seat-priced at $30 throughout, introduces a credit tier within the next two quarters. And the enterprise renewal cycle this autumn, which is the first one where a CIO gets to ask OpenAI why the agent workflow they piloted in Q2 is showing up as a separate consumption line.

The seat isn't dead. It is being quietly demoted.

Footnotes

¶