OpenAI ships spend controls and a cost API for ChatGPT Enterprise

Most enterprise AI budgets are still flying blind. OpenAI shipped new credit usage analytics and granular spend controls for ChatGPT Enterprise on June 18, giving admins a single dashboard for credit burn across ChatGPT and Codex, group-level caps, and a programmatic Cost API.

What the new spend controls actually let admins do

The new tools land in the Global Admin Console, where credit usage from both ChatGPT and Codex is now visible in one place. Admins can see top users, watch credit consumption trends over time, and break spend down by user, product, and model. The same credit data is exposed through a unified Cost API, so finance and FinOps teams can pull it into the same warehouse they already use for cloud or SaaS spend.

On the control side, admins can now set a default credit limit for the whole ChatGPT Enterprise workspace, configure separate limits for specific groups, and create individual overrides for power users. End users can see their credit usage against their budget and request more credits with context, instead of having to ping a manager or file a ticket. OpenAI introduced granular per-role credit limits earlier in 2026, and this release layers on top of that. Employees can include a short note about what they are working on when they ask for more credits, which gives the admin the context they need to decide whether to approve the request.

The pitch is straightforward. Enterprise AI rollouts keep expanding the number of seats, models, and agentic tools that draw from the same credit pool. Without a clear allocation model, a single heavy user running multi-step agent workflows can quietly consume as many credits as a small team of normal users. The new controls give admins a way to set guardrails per group, see who is pushing the limit, and approve extra capacity case by case. The per-user override is the underrated piece, because it lets admins say yes to the heavy user without raising the floor for everyone else in the workspace.

Why enterprise spend became the bottleneck in 2026

The timing is not accidental. Across the industry, enterprises have moved from pilot ChatGPT seats to production rollouts that touch every knowledge worker, and the cost curve has followed. Anthropic, Google, and OpenAI all reported that agentic workloads drive token consumption that compounds with the number of steps a workflow takes. A single agent task can wrap many model calls, and many of those calls use larger models for planning or verification. A 20-step agent run against the flagship model can rack up the same token cost as a few hundred ordinary chat turns, and most admins have no way to see that distinction until the bill arrives.

The same week OpenAI shipped the new spend tools, Microsoft was publicly weighing a fine-tuned, self-hosted version of DeepSeek V4 for its Copilot Cowork product, with the underlying motivation framed as cost management. Reporting cited a comparison that put Anthropic's flagship near $50 per million tokens against roughly $0.87 per million for DeepSeek V4 Pro, a roughly 57x price gap. Even with a model router and a usage-based pricing model, Microsoft is treating per-token economics as a first-class procurement variable. The message to model vendors is clear. Buyers will not accept a flat per-seat price for workloads whose actual cost depends on how aggressively the agent loops.

For OpenAI, the calculus is the opposite. The company does not need to chase a cheaper open-source model to defend its enterprise base. It needs to give admins the same kind of cost control they expect from a hyperscaler cloud bill, with the same kind of allocation model, the same kind of cost-allocation tags, and the same kind of programmatic API. The new Cost API is the most important part of the release for that reason. It signals that OpenAI is now competing on FinOps maturity, not just on model quality. If a CFO can reconcile their ChatGPT bill against the same allocation tags they use for AWS or Snowflake, ChatGPT becomes a line item in the cloud budget instead of a fight at the end of every quarter.

For a typical mid-sized buyer with 5,000 ChatGPT Enterprise seats, the difference between a default credit policy and a thoughtful per-group policy can run into the low six figures per quarter. The new analytics make that gap visible for the first time inside the same console that admins already use to manage users, and the spend controls give them a way to act on the data without re-procuring a separate tool.

Three signals to watch in the next six months

Three signals will tell us whether the new tools actually move enterprise AI spend. The first is whether OpenAI ships native cost-allocation tags for department, project, and cost center. The Cost API is a strong foundation, but finance teams usually want a tag schema that maps cleanly into their existing chart of accounts. The second is whether OpenAI extends the same model to the API platform, where the real money is for many enterprises. Right now the controls cover ChatGPT and Codex inside the workspace, but separate API platform customers do not yet get the same UI. The third signal is whether the per-group limits stay simple, or whether OpenAI adds routing, budgeting, and forecasting primitives over time.

The bigger picture is that enterprise AI is starting to behave like enterprise cloud. Ten years ago, AWS customers learned the hard way that they needed tagging, budgets, and rightsizing before they could hand out accounts. The current wave of AI procurement is going through the same curve in compressed time. OpenAI's June 18 release is the moment the first major model lab built a real cost-control surface, and the rest of the field will have to follow. Anthropic, Google, and the hyperscalers will all have to ship their own versions, because enterprise procurement teams will not standardize on a cost model that only one vendor supports.

For more on how enterprises are buying AI today, see our enterprise AI governance checklist and our look at how Google's Gemini 3.5 Flash is positioning on enterprise AI cost. OpenAI's full announcement is on the company's blog.