Why Can AI Coding Subscriptions Offer So Much Compute for $20?

At first, the economics look strange. Codex is included in the US$20-a-month ChatGPT Plus plan, which OpenAI describes as enough for a few focused coding sessions each week. Its pricing documentation is careful about the boundary: limits vary with the work, additional credits can extend usage, and API-key usage is charged separately by token.^[1]

The gap can still feel substantial. One long agent session can inspect a repository, change many files, run tests, and revise its work. But the apparent API value of that activity is not a disclosure of the provider's marginal cost, and the subscription does not promise an equivalent cash balance of compute.

So why offer a low-priced plan at all? The strongest strategic answer is competition for the AI market itself. The complete answer also needs subscription math, cost improvement, capacity controls, and a route from entry plans to higher-value customers.

The short answer is competition for daily AI habits

The AI platform market is still being formed. OpenAI, Anthropic, Google, GitHub, and others are not merely selling isolated answers. They are competing to become the environment where people write, research, build software, and run parts of a business every day.

That makes an affordable individual plan strategically valuable. Once a developer learns one coding agent, connects repositories, shapes instructions, and builds repeatable workflows around it, the provider has more than a monthly subscriber. It has distribution, product feedback, habit, and a possible path into a team or enterprise account.

Provider announcements in 2026 show the scale and cost of that race. Anthropic said Claude Code's weekly active users had doubled from 1 January and business subscriptions had quadrupled. ^[3]GitHub said agentic usage was becoming the default for Copilot and brought significantly higher compute and inference demand.^[5]

Neither company says, “we price individual plans below cost to lock in users.” That conclusion would overstate the evidence. But rapid adoption, competing entry plans, and large infrastructure investment support a narrower inference: providers have strong reasons to value market share and habit beyond the margin on one month's subscription.

A subscription is not a bag of API tokens

A provider does not need every subscriber to cost less than US$20 every month. It needs the plan to work across the whole customer base. Some people subscribe and use AI lightly. Others run long coding sessions, parallel agents, and automations. The plan is designed around the distribution, not the most expensive person in it.

OpenAI's Codex pricing documentation describes the variance directly: small scripts may consume only a fraction of an allowance, while larger projects and long-running sessions use significantly more. Model choice, context, reasoning, tools, retrieval, and caching all affect usage. The average subscriber and the visible power user are not interchangeable.^[1]

Usage limits then control the expensive tail. Providers can vary limits by plan, model, system load, task type, or time window. They can route suitable work to cheaper models, benefit from cached context, and sell credits when included access runs out. “Included” is generous; it is not unlimited compute at a fixed exchange rate.

Anthropic makes the segmentation visible. Its individual plans listed Claude Pro at US$20 a month, Max 5x at US$100, and Max 20x at US$200 as of May 2026. The entry plan attracts regular users; heavier users reveal themselves and move into higher-capacity tiers.^[2]

Capital gives providers time to build distribution and infrastructure

In February 2026, Anthropic announced a US$30 billion Series G intended to support research, product development, and infrastructure expansion. That is one visible example of the capital available to a leading provider competing for the coding and enterprise AI markets.^[3]

Capital makes aggressive expansion possible. It does not prove that every US$20 account loses money, or that inference is simply being given away. It lets a provider fund capacity, accept longer payback periods, test packaging, and pursue customers whose future value may be much larger than their first subscription.

Compute can work like product-led customer acquisition

For a developer tool, the most persuasive advertisement is useful work completed. Generous included access lets a user feel the product's value, build workflows around it, and show the result to colleagues. In that sense, some inference spending can function like product-led customer acquisition: the product demonstrates itself by doing the work.

Serving costs can fall faster than the headline subscription price

The underlying cost is also moving. Alphabet reported that it reduced Gemini serving unit costs by 78% during 2025 through model optimisation, efficiency, and better utilisation. That figure applies to Gemini, not every provider, but it demonstrates how quickly inference economics can improve even while the familiar consumer price stays fixed.^[4]

Hardware improves. Models become smaller or more efficient. Repeated context can be cached. Schedulers keep expensive accelerators busier. A provider can route a simple task differently from a difficult repository-wide change. These improvements can make today's US$20 plan more capable without requiring the company to absorb the same cost forever.

The generous flat-rate period is already becoming more metered

GitHub supplied unusually direct evidence in April 2026. It said multi-hour agent sessions created much higher inference demand, that GitHub had absorbed much of the escalating cost, and that its premium-request model was no longer sustainable. Copilot moved to included monthly credits with paid usage beyond them from 1 June 2026.^[5]

OpenAI's current Codex pricing follows the same broad structure: Plus includes bounded usage, users can buy credits after reaching a limit, and API-key sessions are charged at standard token rates. The emerging pattern is not the disappearance of subscriptions. It is a subscription that opens the door, followed by clearer metering for sustained agentic work.^[1]

Developers should therefore expect plan structures to keep changing. Providers can adjust included capacity, introduce credits, steer work to different models, or reserve autonomous workloads for higher tiers. A low entry price may remain; the amount and kind of work included at that price need not.

What should developers do while AI access is unusually cheap?

Do not optimise for consuming the largest possible number of tokens. Cheap compute is an opportunity only when it becomes something durable: a working product, a tested automation, a customer insight, a reusable process, or capability you can explain and apply again.

1. Choose a real outcome. Start with one person, one problem, and a result that matters—not a tour of features.
2. Build a complete loop. Use the generous allowance to create the smallest system that can deliver the customer outcome, including testing, failure, recovery, and feedback.
3. Turn output into capability. Inspect, test, explain, improve, and transfer what AI produces. A generated result alone is not learning.
4. Keep the valuable parts portable. Store requirements, decisions, tests, data, and operating knowledge in forms you control so a pricing or model change does not erase the work.
5. Learn the real economics. Track which workflows create value, how much agent work they consume, and whether they would still make sense under metered pricing.

Capture the subsidy without becoming dependent on it

Competition is giving individuals access to AI capability that would have looked extraordinary only a short time ago. Providers are buying distribution and habit; subscription portfolios, usage limits, efficiency gains, and higher tiers help make that strategy economically possible.

The useful response is not to assume US$20 buys unlimited intelligence forever. Use today's access to build assets, products, customer relationships, operating knowledge, and judgment that remain valuable when the plan changes. The compute may be subsidised. The outcome still has to be owned.