AI Moves From Hype to Rate Cards

OpenAI quietly changed Codex on April 2 from a rough per-message system to a token meter, which is the billing method used for application programming interfaces, where every chunk of text in and out gets counted like units on a utility bill. Its help page now lists separate credit charges per 1 million input, cached input, and output tokens for models including GPT-5.4 and GPT-5.3-Codex. (help.openai.com) That is a different kind of AI sale. A per-message price feels like an all-you-can-eat buffet, while token pricing turns coding help into metered electricity where longer prompts, bigger codebases, and more generated output all change the bill. (help.openai.com) OpenAI’s own rate card spells that out with numbers. GPT-5.4 is listed at 62.50 credits per 1 million input tokens and 375 credits per 1 million output tokens, while GPT-5.3-Codex is listed at 43.75 credits for input and 350 credits for output, and Fast mode costs 2 times as many credits. (help.openai.com) The company is also separating customers by buying style, not just by model access. The Codex pricing page sells Plus at $20 a month, Pro at $200 a month, Business with admin controls and single sign-on, and Enterprise with audit logs, role-based access control, data residency controls, and usage monitoring through the Compliance Application Programming Interface. (developers.openai.com) That same page shows how specialized the offers have become. Pro includes access to GPT-5.3-Codex-Spark as a research preview for day-to-day coding tasks, while the Application Programming Interface route is pitched for automation in shared environments like continuous integration, where software gets tested and shipped automatically. (developers.openai.com) OpenAI even gives buyers a back-of-the-envelope budget now. Its help center says Codex costs about $100 to $200 per developer per month on average, with the final number moving around based on model choice, number of running instances, automations, and Fast mode use. (help.openai.com) Meta is moving in the same direction from the model side. VentureBeat reported on April 9 that Meta launched Muse Spark, its first proprietary model since the formation of Meta Superintelligence Labs, and said the model delivers reasoning with more than an order of magnitude less compute than Llama 4 Maverick, Meta’s previous mid-size flagship. (venturebeat.com) That detail matters because compute is the raw material behind the rate card. If a model needs less compute to answer the same coding or reasoning task, the vendor gets more room to package it into cheaper tiers, faster products, or higher margins. (venturebeat.com) Microsoft ran into the legal side of the same shift this week. TechCrunch reported on April 5 that Microsoft’s Copilot terms, last updated on October 24, 2025, still said the assistant was “for entertainment purposes only,” even as Microsoft was trying to sell Copilot into workplaces. (techcrunch.com) Microsoft then said the phrase was legacy language from an earlier Bing Chat era and would be changed in the next update because it no longer reflected how Copilot is used today. When a company is charging for workplace software, old wording stops being a footnote and starts looking like a contract problem. (techcrunch.com, moneycontrol.com) So the center of gravity in artificial intelligence is moving from leaderboard drama to procurement paperwork. Buyers now have to compare token rates, cloud task limits, audit logs, data controls, and terms of use the way they already compare seats, storage, and service-level agreements in ordinary software contracts. (help.openai.com, developers.openai.com, techcrunch.com)