AI pricing shifts to utility logic
Enterprise AI pricing is moving beyond flat seat licences toward hybrid, usage‑based models, with OpenAI publishing Codex credit rates and Anthropic detailing model‑specific charges. (help.openai.com) (platform.claude.com). Reports also say Anthropic may be testing electricity‑style billing, combining fixed commitments with variable consumption. (indiatoday.in)
OpenAI and Anthropic are pricing artificial intelligence less like software seats and more like metered power, with fixed access fees plus charges for what customers actually use. (help.openai.com) OpenAI said on April 2, 2026 that Codex pricing for Plus, Pro, Business, and new Enterprise plans shifted from per-message estimates to token-based billing. Its rate card now lists credits per 1 million input, cached input, and output tokens for each model. (help.openai.com) The new Codex table prices GPT-5.4 at 62.5 credits for 1 million input tokens, 6.25 credits for cached input, and 375 credits for output; GPT-5.4 Mini is lower at 18.75, 1.875, and 113 credits. OpenAI said fast mode uses twice as many credits and estimated average Codex cost at about $100 to $200 per developer per month. (help.openai.com) Anthropic’s pricing page already separates charges by model and token type instead of offering one flat rate. Its current Claude API table lists Claude Sonnet 4.6 at $3 per million input tokens and $15 per million output tokens, Claude Opus 4.6 at $5 and $25, and Claude Haiku 4.5 at $1 and $5. (platform.claude.com) Anthropic also prices repeat work differently. Its pricing page lists cache-hit rates at $0.30 per million tokens for Sonnet 4.6 and $0.50 for Opus 4.6, while batch processing and prompt caching are presented as ways to lower cost for customers who can wait or reuse the same context. (platform.claude.com) This is a change in how enterprise buyers budget for artificial intelligence. A seat license charges for access whether an employee uses the tool heavily or barely touches it; token billing ties more of the bill to how much text, code, and model output actually moves through the system. (help.openai.com) (platform.claude.com) The shift is especially visible in coding agents, which can run long tasks, call tools repeatedly, and consume large amounts of context. OpenAI’s Codex help page says actual credit use depends on the mix of input, cached input, and output tokens in each task, not on a simple count of messages sent. (help.openai.com) Anthropic appears to be pushing the same logic into enterprise contracts. India Today reported on April 15, 2026 that Anthropic has started testing plans with lower per-seat fees, monthly token-use commitments, and the removal of some enterprise API discounts, citing reporting by Let’s Data Science. (indiatoday.in) India Today said some Anthropic roles were repriced to $20 a month from earlier ranges of $40 to $200, but customers would also commit to estimated monthly token usage even if actual use came in lower. Anthropic’s own public pricing pages do not spell out those reported enterprise contract terms, so that part of the picture remains based on outside reporting. (indiatoday.in) (claude.com) Sam Altman has described the destination in utility terms. India Today reported that at the BlackRock US Infrastructure Summit in Washington, District of Columbia, the OpenAI chief executive said artificial intelligence could become a daily service people pay for by usage, like electricity. (indiatoday.in) The result is not the end of subscriptions but a hybrid model: access still has a monthly price, while heavy workloads add a metered bill on top. OpenAI’s ChatGPT Business page says customers get “the flexibility to add credits as needed,” which is the language of utility pricing moving into office software. (chatgpt.com)
