OpenAI moved Codex, its agentic coding tool inside ChatGPT, from fixed per-seat pricing to usage-based, token-metered billing on April 2. The stated goal was flexibility: teams can now bring Codex into ChatGPT Business and Enterprise plans without committing to fixed seat costs up front. The mechanics underneath are what finance and IT leads actually need to model before rolling this out further.
Bring Codex to your team without fixed seat costs.
— OpenAI Developers (@OpenAIDevs) April 2, 2026
We’re rolling out usage-based pricing for Codex in ChatGPT Business and Enterprise plans, so teams have a more flexible way to get started. pic.twitter.com/dyfuK5ouXS
What is actually new
Codex billing now runs on credits per million tokens, tiered by model: GPT-5.5 costs 125 credits per million input tokens, 12.5 for cached input, and 750 for output; GPT-5.4 runs roughly half that; the coding-specific GPT-5.3-Codex model sits lower still. OpenAI says a typical Codex task consumes 5 to 45 credits depending on model and task complexity, and independent estimates put real-world spend around $100 to $200 per active developer per month. The rollout expanded to existing Enterprise, Edu, Health, Government, and Teachers plans by April 23.
Eligible workspaces received promotional credits, $100 per new Codex-only seat up to $500 total, to soften the transition. Base ChatGPT plan usage was not affected; the change applies to purchased Codex credits and Codex-only seats specifically.
What this means for IT and finance budgets
The detail most coverage of the April announcement skipped: starting June 24, 2026, workspaces without a prior Codex seat can no longer add new ones under the old seat-based structure. Every new ChatGPT Business or Enterprise customer bringing on Codex now goes straight to usage-based billing, full stop. That is a meaningful shift in how this gets budgeted. Per-seat pricing is a headcount problem finance teams already know how to forecast. Token-based billing is a consumption problem, closer to cloud infrastructure spend, and it needs the same guardrails: usage caps, workspace credit pools, and alerts before a heavy sprint quietly triples a monthly bill.
It also reframes the vendor comparison we laid out when covering Cursor’s Composer 2.5 pricing against Opus 4.7 and GPT-5.5: OpenAI is now pricing Codex access the same way it prices raw API tokens, which makes side-by-side cost comparisons between coding tools more direct than they used to be, but also means the cheapest sticker price on a Business plan is no longer the whole story.
What to watch
Watch whether Anthropic and Google follow the same path for their own Business-tier coding tools. Usage-based pricing is easier for a vendor to defend as usage scales, and harder for a buyer to forecast. The same tension is already playing out in how vendors price agent fleets generally, the subject of our look at what 100 autonomous coding agents actually cost per seat.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
