On June 24, 2026, OpenAI quietly removed Codex seats from all new ChatGPT Business plans. If your workspace did not already have a Codex seat before that date, you cannot add one at a fixed monthly price anymore. The only path forward is usage-based billing. Two days earlier, MindStudio published an analysis arguing that the entire $20-per-month AI subscription model is unsustainable. Neither event made mainstream headlines. Both are signals that the pricing structure most small businesses rely on for AI tools is about to change underneath them.
The economics never actually worked
The flat-rate model for AI was always a user acquisition play, not a sustainable business. OpenAI has publicly acknowledged that heavy users of the $200-per-month ChatGPT Pro plan can still cost the company money. At $20 per month, the math is worse. Every message you send to GPT-4o or Claude spins up compute clusters, processes tokens, and generates a response. A casual user who sends twenty messages a day is cheap to subsidize. A power user running multi-step research sessions, generating images, and using AI to draft documents across dozens of daily sessions is not.
The shift to agentic AI makes this problem dramatically worse. An AI agent does not just answer one question. It browses web pages, writes and executes code, calls APIs, generates images, loops back to check its own work, and spawns sub-agents for parallel tasks. A single agent run can involve hundreds of model calls. If you are running an AI agent that monitors your inbox, summarizes threads, drafts responses, and updates your CRM across hundreds of emails per day, the compute bill looks nothing like someone asking ChatGPT for a recipe. Flat subscription pricing was designed for conversations. It was never designed for autonomous background workloads at scale. That world arrived in 2026.
OpenAI just proved the point
The June 24 change to ChatGPT Business is not an isolated decision. It is the first visible crack in a pricing model that every major AI provider is rethinking. OpenAI's own help center now describes two seat types for Business: standard ChatGPT seats at a fixed per-user monthly cost, and Codex seats that are purely usage-based. Existing workspaces that had Codex seats before June 24 can keep managing them. New workspaces cannot get them at all. The direction is unmistakable. Fixed-price access to compute-intensive features is being replaced by metered billing, and the transition is happening now, not in some hypothetical future.
This matters for small businesses because ChatGPT Business is the plan most teams of five to twenty people are on. If you added Codex seats in May, you are grandfathered. If you are signing up today, you are paying per use. The same pattern will likely extend to other features over the next twelve months. Image generation, advanced data analysis, and extended context windows are all compute-heavy features that providers currently subsidize. When those move to usage-based pricing, the monthly bill that used to be predictable will become variable, and the businesses that do not budget for it will get a unpleasant surprise.
The measurement gap makes this worse
A QuickBooks AI Impact Report published in May 2026 surveyed 34,000 small businesses in partnership with the University of Chicago. The headline numbers look great: 77% of respondents reported regular AI use, up from 48% in mid-2024, and 41% reported revenue increases attributed to AI. But when researchers asked how those businesses were measuring the improvements, the answers fell apart. More than 50% described a general feeling that their business was doing better. Less than half tracked specific metrics. The productivity number was based on self-reporting, not time studies. The revenue attribution was correlation, not controlled measurement.
This is the trap. Businesses are spending money on AI and believing it is working without measuring whether it actually is. As pricing shifts from flat-rate to usage-based, that vague sense of value becomes a real financial risk. If you do not know what your AI tools actually save you in hours or dollars, you cannot make informed decisions about what to keep, what to cut, and what to double down on when the bill changes.
What to actually do right now
Start tracking. Not next quarter, this week. Pick the three AI tools your business uses most. For each one, log the monthly cost, the number of people using it, and one concrete output metric. For a writing tool, that might be blog posts or email drafts produced per week. For a customer service tool, tickets resolved per hour. For a coding assistant, pull requests merged. You do not need a sophisticated dashboard. A spreadsheet with four columns is enough to start.
Next, audit your seat count. The Goldman Sachs small business AI survey from March 2026 found that 93% of businesses using AI reported positive impact, but only 14% had fully integrated it into core operations. That gap means most businesses are paying for seats that get used occasionally, not daily. Every unused seat on a fixed plan is wasted money. Every unused seat on a usage-based plan is at least not generating charges, but it still represents an investment in access nobody is using.
Finally, start thinking in terms of cost per outcome rather than cost per seat. The old model was simple: $20 per person per month, use it or do not. The new model will charge you for what you actually consume. That is better economics for the providers and, if you are paying attention, better economics for you too. A business that spends $200 per month on AI and can point to ten hours of saved labor has a clear return. A business that spends $200 per month and says "it feels like it helps" is flying blind.
The $20-per-month AI subscription is not going to vanish overnight. But the trend is clear, and the OpenAI change this week is the loudest signal yet. The businesses that start measuring now will make smooth transitions when their pricing changes. The ones that do not will be arguing with their accountant about a line item they never tracked.