On published prices, GitHub Copilot CLI is the cheapest of the three terminal agents and Claude Code is the most expensive. A Copilot Business seat is $19 a month and includes 1,900 AI credits, which is $19 of model usage at a cent each. A Codex CLI seat rides on a ChatGPT Business plan at $20 per user billed annually. Anthropic publishes real deployment numbers for Claude Code: around $13 per developer per active day and $150 to $250 per developer per month. For ten developers that is roughly $190, roughly $200, and $1,500 to $2,500. Those three numbers buy genuinely different things, and the rest of this piece is about what.
Every figure here comes from the vendor's own documentation, read in August 2026. Where we do arithmetic on top of a published rate, the assumption is stated in the sentence above the table.
The three terminal agents, priced side by side
All three are the same shape of product: an agent that lives in your terminal, reads your files, runs commands, and edits code in a loop. They meter you in three completely different units, which is why price comparisons of these tools usually go wrong.
| GitHub Copilot CLI | Codex CLI | Claude Code | |
|---|---|---|---|
| Entry price | $0, included on Copilot Free | $0, included on ChatGPT Free | Included on a Claude plan, or per token on the API |
| Typical team seat | Business $19 per seat | Business $20 per user billed annually, $25 monthly | Team or Enterprise seat, or metered API |
| Metering unit | AI credits, 1 credit = $0.01, priced per model | Credits per message plus a rolling five hour message cap | A per-seat allowance on a rolling five hour and weekly window |
| Included usage | 1,900 credits per Business seat, pooled across the org | 25 to 200 local GPT-5.6 Terra messages per five hour window on Plus | Anthropic states the seat allowance is the default ceiling |
| Overage | $0.01 per credit, the same rate as included credits | Buy a higher plan, or use an API key at standard rates | Turn on usage credits and set a spend limit |
The first row is the one people misread. Copilot CLI being free on Copilot Free does not mean the CLI is free, it means the CLI is not the thing being sold. GitHub states plainly that all plans include Copilot CLI. What your plan buys is the credit balance the CLI spends, and since code completions and next edit suggestions are never billed, agent work like the CLI is the entire variable part of a Copilot invoice.
What one terminal session actually costs
GitHub publishes per-million-token rates for every model it offers, so a session has a computable price. Our assumption, and it is ours rather than GitHub's: one session is 300,000 input tokens and 30,000 output tokens, which is what a multi-turn agent run looks like when the whole conversation is resent with every tool result. Adjust it up if your repository is large.
| Model in Copilot CLI | Rate per 1M in / out | Credits per session | Sessions on a $19 Business seat |
|---|---|---|---|
| GPT-5.6 Luna | $0.20 / $1.20 | 9.6 | 197 |
| GPT-5 mini | $0.25 / $2.00 | 13.5 | 140 |
| Claude Haiku 4.5 | $1.00 / $5.00 | 45 | 42 |
| GPT-5.3-Codex | $1.75 / $14.00 | 94.5 | 20 |
| GPT-5.4 (default) | $2.50 / $15.00 | 120 | 15 |
| Claude Sonnet 4.6 | $3.00 / $15.00 | 135 | 14 |
| Claude Opus 5 | $5.00 / $25.00 | 225 | 8 |
Read the top row against the bottom one. Same seat, same work, and the model dropdown decides whether a developer gets 197 sessions a month or 8. That is the single biggest lever anyone has over a Copilot bill, and it is a setting rather than a purchase. We take the whole table apart, including Copilot Pro and Max, on our GitHub Copilot CLI pricing page.
Copilot CLI vs Claude Code: which is cheaper?
Copilot CLI, by a wide margin on published numbers, and it is not close. A $19 Copilot Business seat includes $19 of model usage and charges overage at exactly the included rate, so there is no penalty for going over. Anthropic's own cost documentation reports around $13 per developer per active day and $150 to $250 per developer per month across enterprise deployments, staying under $30 per active day for 90% of users. Nobody is hiding anything here. Anthropic publishes those numbers because Claude Code is doing more work per session.
Here is the comparison that matters for a ten person team. Say each developer wants 30 real agent sessions a month on a strong model, Claude Sonnet 4.6 inside Copilot CLI. That is 300 sessions at 135 credits, or 40,500 credits. Ten Business seats include 19,000, so you buy 21,500 credits of overage at a cent each, which is $215. Total bill: $190 in seats plus $215 in usage, or $405 a month, about $40.50 per developer. The same ten developers on Claude Code cost $1,500 to $2,500 a month at Anthropic's published range.
That is a four to six times difference, and it is real. What it does not tell you is whether 30 Copilot sessions get the same work done as 30 Claude Code sessions. In our experience they do not, and the gap widens on long, multi-file tasks. Price the outcome, not the session.
Codex CLI and the five hour window nobody budgets for
Codex is the odd one out because it does not meter you against a balance you can watch drain. It meters you against a rolling five hour window of messages. On a Plus plan at $20 a month, OpenAI publishes 25 to 200 local GPT-5.6 Terra messages per five hour window, and 250 to 2,000 on the cheaper Luna model. Pro at $200 gives 20x limits. Business is $20 per user billed annually or $25 monthly.
The practical difference is what failure looks like. A Copilot developer who overspends gets an invoice. A Codex developer who overspends gets locked out at 2pm on a Tuesday, on the same plan that was fine yesterday, because the window is a function of how hard the morning went. Neither is worse in the abstract. One is a finance problem and one is a scheduling problem, and different organizations are better at absorbing one than the other.
Where each one hides its cost surprise
- Copilot CLI: the model dropdown, worth up to a 47 times difference in the cost of one session. Also worth knowing: credits do not roll over and reset at 00:00:00 UTC on the first of the month, so a migration you want an agent to grind through should live inside one calendar month.
- Codex CLI: image work. OpenAI notes that image tasks consume limits three to five times faster on average than text, so a design-heavy sprint burns the window at a rate the plan page does not suggest.
- Claude Code: long sessions. Anthropic warns that the full conversation is resent with every request, so a one line question in a session left open all day still draws usage for the entire thread. The fix is free and takes one keystroke: clear between unrelated tasks.
There is a fourth surprise that spans all three, and it lands on a different budget line than the one you are watching. Any of these tools can be pointed at your own provider, and when Claude Code runs through Amazon Bedrock, Google Cloud or Microsoft Foundry, that spend appears in your cloud bill rather than on a SaaS invoice. Teams that have started routing agents this way usually find the only way to see the real total is to pull cloud and SaaS spend into one view, because otherwise engineering sees the seat fees and finance sees the inference and nobody sees the sum.
Is GitHub Copilot CLI free?
It is included on every Copilot plan, including Copilot Free, so there is no separate purchase to make. It is not unlimited. Every interaction spends AI credits from your plan allowance, and Copilot Free restricts you to automatic model selection, which removes the only real lever you have over what a session costs. Evaluate on Free, but price on Pro or Business.
Which terminal agent should a US engineering team buy?
If your team already pays for Copilot seats, start with Copilot CLI and change nothing else. It costs nothing extra, needs no second vendor and no second security review, and the credit table above tells you your ceiling before you commit. Standardize on a cheap model for routine work and reach for an expensive one deliberately.
If your team lives on ChatGPT already, Codex CLI is the same argument in the other direction, with a hard window instead of a balance. And if you have tried both and keep hitting the ceiling of what a cheaper agent can finish, Claude Code is priced the way it is because it does more per session. That is a capability decision with a real number attached, and Anthropic publishes the number, which is more than most vendors do.
The question none of the three answers
All three of these put an agent in your terminal, and all three need you sitting in front of it. You approve steps, you read output, you keep the session alive, and every minute of that attention is billed to you in credits, messages, or allowance. That is a real product and a good one. It is also not the same thing as handing off a task.
Agentcode takes a task against a connected GitHub or GitLab repository, plans it, edits the code, runs your test suite, and opens a pull request for a human to read. It never merges on its own, so your branch protection and your reviewers stay exactly as they are, and it costs $29 a month billed yearly flat. No credit balance, no five hour window, no model dropdown deciding your invoice. If you want the Copilot side of this in more depth first, the Copilot agent pricing breakdown covers agent mode and the cloud agent, and Codex against Claude Code takes those two apart properly.