All You Can Eat Economics for AI Coders
Lunch tastes better when you know how much it really costs.
TL;DR Not only you should be careful with closed AI coding tools (consider avoiding them, there are good open alternatives) but you should also be suspicious of fixed rate subscriptions for using such tools.
TANSTAAFL; all-you-can-eat deals are only profitable under one of the following conditions:
Predictable Maximum Consumption. For example: actual all-you-can-eat buffets. Because there’s a limit to how much even the hugriest customers can eat — 2-3 servings at most, and probably more like 1.5 servings on average — you only need to set the price high enough to cover that and add some margin.
Zero Marginal Cost. For example: digital content, like unlimited access to a digital library. The content has already been written and the cost of allowing another reader to access it is approximately zero, so once you’ve covered the initial cost of producing the content, it’s all profit.
Loss Leader. The seller is losing money on each new customer, but has a plan for offsetting this later — by upselling another, profitable deal, or by locking the buyer in and hiking the price later until it reaches profitability. For example: girls night at the disco. Instead of charging entrance, girls get in for free. If that was that, the disco would lose money, but more girls means more boys and more expensive drinks, so overall the night is profitable.
Fake News. Consumption is actully capped, to make it predictable, but the information on the cap is in the small print, tempting buyers to sign up with the hope of getting unlimited supply and without noticing the real limits. For example: unlimited cloud storage (subject to a “fair use” policy that allows the provider to throttle arbitrarily).
What about fixed rate AI coding subscriptions, like the “MAX” subscriptions from Claude, Cursor, or ChatGPT? Consumption is unpredictable, since agents can work autonomously and a user can run multiple agents in parallel. The marginal cost is far above zero - every token inferred for the user is uniquely generated for them on expensive hardware.
All-you-can-code subscriptions are some combination of (3) and (4). All practice some form of throttling, often quite opaque and difficult to predict. All are, to an extent, loss leaders. The providers understand that the more you use their tool now, the more likely you are to stick with it for the long run and even recommend it to others. The goal is not short-term profitability, but market-share expansion and brand recognition. Once that’s achieved, it’s time to rework the product offering and make it profitable.
There’s nothing wrong with that — these are all standard business practices. But if you, the buyer, are aware, you can make wiser choices and avoid being the victim of an exploitative practice.
The same functionality is available from tools and service providers that are more transparent and charge per use (in some cases with modest discounts on bulk purchase and even limited but generous free allowance).
You should consider adopting such tools. Lunch tastes better when you know how much it really costs.
Aider, Codex CLI, GitHub Copilot, Cline, OpenHands, OpenCode, Roo Code, and others support models from any of the providers, as well as local open open models. You bring your own API keys, or pay per use through the provider.
Both OpenAI and Google Deep Mind have a generous free tier for using their SOTA modals like Gemini 2.5 Pro and Flash, GPT-4.1, codex-mini, and o3 (in both cases you need to be OK with them looking at your requests and potentially using them in research and training).
Gemini CLI is a special case where the tool is open-source and the economic arrangements are straightforward, but only the Gemini models are supported in the mainline version — the Gemini team confirmed that they intend to keep it that way for now. But nothing in the licensing terms prevents the community from forking and creating an “Open Gemini” version that works with other providers.