Pricing UX shows hidden caps

What looks like a full AI quota often is not fully usable in production, because billing meters and capacity caps sit on top of each other. OpenAI, Anthropic, and Google all publish separate limits for requests, tokens, projects, regions, or monthly spend that can stop usage before a dashboard “100% available” figure is exhausted. (developers.openai.com) (platform.claude.com) (docs.cloud.google.com) The basic split is simple: pricing tells you what one unit costs, while quotas tell you how fast or how far you can consume those units. OpenAI says its API can be limited by requests per minute, requests per day, tokens per minute, tokens per day, image calls per minute, and monthly usage limits, with whichever cap is hit first ending the run. (developers.openai.com) Anthropic uses the same layered structure. Its Claude API docs say organizations face both monthly spend limits and rate limits, and that short bursts can trigger errors even when the headline per-minute allowance looks sufficient on paper. (platform.claude.com) Google’s Vertex AI adds another layer by tying some quotas to project and region, not just account-wide usage. Its generative AI docs list separate per-region and per-model quotas, and say tuned model inference can share the same quota as the base model instead of getting a fresh pool. (docs.cloud.google.com) That is how a customer can buy “capacity” and still find only part of it practical to use in real workloads. A long prompt can consume token-per-minute headroom before request count is reached, a bursty workflow can hit per-second enforcement inside a per-minute plan, and a shared org limit can let one team starve another. (developers.openai.com) (platform.claude.com) The pricing pages make the mismatch easier to miss because they foreground unit economics, not delivery constraints. OpenAI’s pricing page leads with per‑million‑token rates, Batch discounts, and paid service tiers, while Google’s Vertex AI pricing page lists token prices and separate charges for grounded search queries that can stack on top of model usage. (openai.com) (cloud.google.com) Those extra meters change what “usable quota” means. Google says grounded Gemini requests include 5,000 search queries per month at no charge and then bill additional queries at $14 per 1,000, while OpenAI prices web search at $10 per 1,000 calls and Batch processing at a 50% discount from standard rates. (cloud.google.com) (openai.com) Service firms are changing their own pricing partly because artificial intelligence compresses labor time without removing infrastructure costs. WPP said in March 2026 that it is pushing compensation tied to sales and brand outcomes rather than hours worked, as AI tools reduce the time needed to generate large volumes of creative work. (storyboard18.com) That shift moves the commercial argument away from “how many hours were used” to “what entitlement was actually delivered.” When compute, token throughput, search calls, and org-wide caps all shape delivery, a clean hourly bill or a simple quota bar no longer describes what a buyer can reliably consume. (developers.openai.com) (cloud.google.com) (storyboard18.com) The practical fix is not a new buzzword but clearer product design: show the hard cap, the burst cap, the shared cap, and the billable add-ons in the same place. Until vendors do that, “100% available” will keep meaning less than it appears to mean. (developers.openai.com) (platform.claude.com)