OpenAI Codex billing spikes reported

OpenAI’s Codex pricing changed in a way that looks cheaper at first glance and riskier once you read the fine print. The company cut ChatGPT Business seats from $25 to $20 and added Codex-only seats with no fixed monthly fee. But starting April 2, 2026, Codex for Business and Enterprise also moved onto token-metered credit billing, with Enterprise migrations finishing April 23. That means coding work that used to feel bundled now maps directly to usage — and for some teams, that is where the bill shock anxiety starts. ### What actually changed? OpenAI split team billing into two seat types. A standard ChatGPT seat still has a fixed per-user price and includes baseline Codex access. A Codex seat is different — no fixed seat fee, Codex only, and all activity draws from workspace credits. For Enterprise and Edu, shared credit pools already mattered, but now Codex pricing itself is explicitly tied to token use rather than rough per-message estimates. (openai.com) ### Why are people calling this a billing spike? Because “usage-based” sounds neat until you realize the meter is no longer hidden. OpenAI’s own help pages say the old system was replaced with direct pricing for input, cached input, and output tokens. So a long coding session, a big repo, repeated cloud tasks, or a model that emits lots of output can burn through credits much faster than a team expected from the old bundled mental model. The complaint is less “OpenAI added billing” than “the same behavior now shows up as a much larger visible number.” (help.openai.com) ### How does the meter work? The rate card is the key. GPT-5.5 costs 125 credits per 1 million input tokens, 12.5 for cached input, and 750 for output. GPT-5.4 is half that. GPT-5.3-Codex and GPT-5.2 sit at 43.75 input and 350 output. Fast mode costs more on supported models, and OpenAI notes that code review uses GPT-5.3-Codex. So the expensive part is often not the prompt going in — it’s the model’s output and the number of times a task reruns. (help.openai.com) ### Why does output matter so much? Because coding agents are chatty. They inspect files, plan steps, generate patches, explain changes, retry failed tasks, and sometimes fan out across cloud workflows. Think of it less like sending one message and more like hiring a junior engineer who narrates every step and rewrites code three times. The meter follows all of that work. OpenAI even says larger codebases, long-running tasks, and extended sessions use significantly more per message. (help.openai.com) ### Is this only a Business problem? No — but Business and Enterprise feel it differently. Business users still have per-seat limits on advanced features, then dip into shared credits if the workspace buys them. Enterprise workspaces use a shared credit pool at the contract level, and admins can allow overages or additional purchases. OpenAI also says owners get threshold alerts and can use role-based controls to manage spend, which is basically the safety valve if finance is worried. (developers.openai.com) ### Did OpenAI frame this as a price cut? Basically, yes. The company pitched the change as easier adoption and clearer cost tracking. It highlighted $20 Business pricing, promo credits for eligible workspaces, more than 2 million weekly Codex users, and 6x growth in Business and Enterprise Codex usage since January. That framing makes sense if you want pilots to start small. But it also means the customer, not OpenAI, now carries more of the variability. (help.openai.com) ### So what’s the real issue? The real issue is budgeting. Transparent pricing is good. Unbounded pricing inside a tool employees can hammer all day is harder. If a workspace enables auto-recharge, overages, fast mode, and top-tier models without controls, a “cheap” seat can turn into a very expensive month. And OpenAI’s own documentation says average Codex cost is about $100 to $200 per developer per month, with large variance. That “large variance” is the whole story. (openai.com) ### Bottom line This is not just a price increase story. It is a pricing-model story. OpenAI made Codex easier to start and easier to meter, but also easier to overspend on if teams treat agentic coding like flat-rate SaaS instead of metered compute. (openai.com) (help.openai.com)