How much do AI coding tools really cost?

The short answer is that you can start for nothing and do real work for $10 to $20 a month. As of October 2026, GitHub Copilot Pro costs $10 a month, Claude Pro (which includes Claude Code) costs $20 a month or $17 a month billed yearly, ChatGPT Plus (which includes Codex) costs $20, and Kiro Pro costs $20. Higher tiers, for people who run agents for hours each day, sit between $100 and $200, and a few go higher.

The longer answer is that two $20 plans can give you very different amounts of work. Each company measures usage its own way, and agents, which read many files and run long tasks, consume far more than a quick question or a line of autocomplete. Prices below are in US dollars, before tax, and change often.

The main plans at a glance

Codex's details come from OpenAI's Codex pricing page, and Lovable's from its pricing FAQ.

A worked example: one month of a learner's coding

Suppose you spend about an hour a day on a side project: some autocomplete, a few chat questions, and two or three agent tasks a week, such as "add a search box and its tests."

• On Copilot Pro ($10): autocomplete is free of credits. Chat and agent work draw from the included $15 of credits; GitHub says a quick question to a small model costs a fraction of a credit, while a long agent session on a large model costs much more. If you run out, you can set a budget, for example $10 for another 1,000 credits, or wait for next month.
• On Kiro Pro ($20): your 1,000 credits cover prompts, spec writing and agent "hooks." Simple prompts can cost less than one credit, spec tasks usually cost more. Running out mid-month and buying 250 extra credits would add $10. Unused credits do not carry over.
• On Claude Pro ($20): there is no credit counter. You work until you reach the limit for the current five-hour window, then wait for it to reset; a weekly cap sits on top. Your chats in Claude and your Claude Code sessions draw from the same pool, so a long research chat leaves less room for coding that afternoon.

The cheapest plan is the one that matches your rhythm. Short daily sessions suit time windows; occasional bursts suit credits you can top up.

What do credits, tokens and usage windows mean for your monthly bill?

Tokens are the basic unit underneath everything. A model reads and writes text in small pieces, roughly three or four characters of English each. Every file an agent opens, every message in the conversation and every line it writes is counted in tokens. This is why agents cost more: to fix one bug, an agent may read dozens of files and loop through several attempts, which can add up to many thousands of tokens. Tools that let you bring your own API key bill these tokens directly at the model maker's price.

Credits are a tool's own currency layered on top of tokens. GitHub ties its credits to dollars (one credit is one cent). Kiro and Lovable use credits as units of work whose cost depends on the model and the size of the task. Credits make bills predictable up to a point: you know what you paid for, but not exactly how many tasks it buys, because a bigger model or a longer task burns credits faster.

Usage windows replace a counter with a time limit. Claude's plans reset on a rolling five-hour window, with weekly limits on paid plans, and OpenAI estimates Codex usage per five-hour period on Plus. You don't get a surprise bill, but you can be stopped in the middle of an afternoon. Both companies let you continue with pay-as-you-go credits or an API key if you need to.

Three habits keep the bill under control. Use a smaller, cheaper model for routine tasks. Keep tasks small and specific, so the agent reads fewer files. And check the usage page weekly for the first month, before you decide whether you need a bigger plan.