What Coding Agents Actually Charge You in 2026
Claude Code, Codex, Copilot, Cursor, Gemini CLI, and Grok Build meter your work in four incompatible ways. The meter, not the model, is what decides your bill.
Every comparison of coding agents ranks the models, and every one of those rankings is stale within a quarter. The part that stays true longer is the billing model, because changing it annoys customers and vendors do it rarely.
There are four meters in this market right now, and which one you are standing on explains almost everything about how your month goes: an opaque usage multiplier attached to a seat, dollar denominated credits, a free request quota, and raw tokens.
All prices below came off vendor pricing pages on 21 August 2026. This category reprices often, so treat the numbers as a snapshot and the structure as the durable part.
The four meters
Claude Code, Codex, and Cursor all sell a seat with a usage allowance that is described in multiples rather than units. Claude is $20 a month on Pro ($17 on annual billing), with Max starting at $100 for five times Pro’s usage and a twenty times tier above that. Codex follows the same ladder almost exactly: $8 on Go, $20 on Plus, then $100 for five times and $200 for twenty times, with per model message counts inside a rolling five hour window. Cursor’s Hobby tier is free, Pro is $20, and Pro+ and Ultra are sold as three times and twenty times Pro’s agent limits, with Teams at $40 per user.
Copilot is the odd one out, and it is the reason people started paying attention to this at all. Its plans are denominated in dollars: Pro is $10 a month and includes $15 of credits, Pro+ is $39 with $70, Max is $100 with $200, and one credit is one cent. Code completions do not draw down credits. Chat, agent runs, and CLI work do, and overage is purchased at the same cent per credit.
Gemini CLI is the only free option that is free in a way you can plan around. It is Apache 2.0 and a personal Google account gets 60 requests a minute and 1,000 a day at no cost. An API key or Vertex moves you onto usage billing when you outgrow that.
Grok Build is the token play. xAI lists grok-build-0.1 at $1.00 per million input tokens and $2.00 per million output under 200k context, with a 256k window. That is the cheapest per token number among the coding specific models, and it is the only one of these where your spend is a straight function of how much text moves.
Why the credit model felt like a price increase
When Copilot moved its agent and chat usage onto credits, nothing about the underlying compute changed. What changed is that the bill started itemizing.
A seat with an opaque multiplier hides variance. Two developers on the same $20 plan can consume wildly different amounts of compute, and neither one sees a number until one of them hits a limit and gets told to wait. Under a credit balance, that same variance shows up as a falling dollar figure in a dashboard, which reads as a new charge even when it is the same work.
Both models have a real failure mode. Opaque multipliers fail by stopping you mid task with no way to have predicted it, and no way to pay to continue except upgrading a tier. Credits fail by not stopping you at all, which is how a $10 plan turns into a several hundred dollar month if nobody set a spending cap. If you are on Copilot, the cap and the alert thresholds are the setting that matters, and they are worth checking before the next agent run rather than after.
What the terminal tier changed
The CLI agents are a different product from the editor agents, even when the same company sells both. An agent in an editor assumes a human is present for every diff. An agent in a shell assumes it might be running under tmux, over ssh, or inside CI with nobody watching, which is why plan modes, approval prompts, and headless flags are the features that separate them rather than autocomplete quality.
That difference is also where the money leaks. Interactive use is naturally rate limited by how fast a person can read. An unattended run is limited only by the task, and a loop that retries a failing test twenty times costs twenty times as much on a token meter, one large chunk of a daily quota on Gemini CLI, and possibly nothing visible at all on a seat plan until the seat stops working.
Before wiring any of these into automation, find the flag that caps a run. Every one of them has some version of it, and the default is never the value you want in CI.
Picking a meter rather than a tool
If you want a predictable invoice and can tolerate an occasional wall, buy a seat. Claude Pro, ChatGPT Plus, and Cursor Pro are all $20 and all sell you the same promise: a fixed monthly number in exchange for an allowance you cannot precisely measure. The $100 and $200 tiers exist because heavy users hit that wall daily, and moving up is the only lever those plans give you.
If you want visibility more than predictability, Copilot’s credits are the most honest pricing here, and the one that lets a manager answer the question “what did agent work cost us in July” without guessing. Set the spending limit first.
If your usage is spiky and mostly small, Gemini CLI’s free quota covers a real working day, and it costs nothing to have installed alongside whatever else you use.
If you are running agents unattended at volume, tokens are the only meter that scales linearly and the only one where optimization pays back. That is the argument for Grok Build’s rate, and equally the argument for pointing any CLI agent at an API key rather than a subscription once the seat stops making sense.
Most teams end up with two of these, not one, and that is a reasonable outcome rather than indecision. The mistake is running a second one for months without checking which meter absorbed the work.
Open your billing page for whichever agent you used most last month and find the actual number. On a seat plan, the number that matters is how many times you hit the limit, not the invoice, because that is the one telling you whether you are on the wrong tier.