Claude Code, GitHub Copilot or Codex for a UK development team: what each costs and does.

Copilot lives in the editor and in GitHub: inline completions, next-edit suggestions, chat about the file you have open, pull-request review and summaries. Its value is continuous and small: a few seconds saved on every line, and a reviewer that never gets tired. Claude Code is an agent: you describe a change, it reads the codebase, plans, edits across files, runs the tests, and reports back, in the terminal or from your editor. Its value is lumpy and large: a migration, a feature, a refactor that would have taken an afternoon. Codex is OpenAI’s version of the same agent, in the terminal, the browser and inside ChatGPT, with cloud tasks that run while you do something else.

Teams that compare them as if they were substitutes end up disappointed with whichever they chose. The useful comparison is Claude Code against Codex for the agent seat, with Copilot as the editor layer under either.

Claude Code is not a product you buy separately. It ships with Claude Pro ($17 or $20 a month), Max ($100 or $200), and every Team seat ($20 or $25 Standard, $100 or $125 Premium), drawing on the same usage allowance as chat, metered in rolling five-hour windows with weekly limits on paid plans. A developer who uses it all day will exhaust a Standard seat and belongs on a Premium seat or Max; Anthropic’s own documentation puts typical enterprise spend at a few pounds per developer per active day. Enterprise is $20 a seat plus usage at API rates.

GitHub Copilot moved to usage-based billing on 1 June 2026. Business stays at $19 a user a month and Enterprise at $39, but what the fee buys is now a monthly allowance of AI credits, $19 and $39 respectively, spent per interaction at rates that vary by model; code completions and next-edit suggestions stay unlimited on every paid plan. Individual plans are Pro at $10, Pro+ at $39 and Max at $100, each with a credit allowance above the base. Enterprise requires GitHub Enterprise Cloud on top.

Codex is included with ChatGPT: Plus at $20 with message limits per five-hour window that vary by model, Pro at $100 or $200 with five or twenty times Plus, Business at $20 a seat on credits, Enterprise on a credit pool. A team already on ChatGPT Business has Codex already.

List prices on 25 September 2026, converted at USD 1 = GBP 0.7545. Heavy agent use on any plan spends credits or hits limits faster than chat.

GitHub states that Copilot Business and Enterprise data is not used to train its models, and that interactions on the Free, Pro, Pro+ and Max individual plans may be used for training unless the user opts out. Claude Code inherits the plan it runs on: from a Team or Enterprise seat or the API it is covered by Anthropic’s commercial terms and not used for training; from a personal Pro or Max account it falls under the consumer training choice, which is presented switched on. Codex inherits ChatGPT’s position: Business and Enterprise workspaces are not trained on; Plus and Free are, unless the user turns it off.

The rule for a UK firm is the same in all three cases: developers work from a business seat, not a personal one, so the firm’s repositories are covered by a contract rather than by an employee’s settings page. A developer on a personal Max plan working on client code is the most common way a firm breaks this without noticing.

This is the choice that matters, and it is closer than either vendor’s marketing suggests. Our own engineering runs on Claude Code: it is what built and maintains the two products we operate, and it is what we reach for on client work. The reasons are practical rather than ideological. It holds a large codebase in context, it is disciplined about running tests before claiming a change works, and it is honest when it cannot do something. Codex has cloud tasks that run in parallel while you do other work, tighter integration for a team already living in ChatGPT, and is a fair choice for a team whose developers prefer it.

The test that settles it is the one we run with clients: two weeks, the same five real tickets from your backlog given to each tool by the same developers, scored on whether the change merged without rework and how long the developer spent steering. Buy on that, not on a benchmark from a vendor blog.

GitHub Copilot Business for all five, because completions and pull-request review inside GitHub are cheap and used all day: about £72 a month. Claude Team with three Standard seats and two Premium seats for the people running agents all day: about £195 a month. Total under £270 a month, about £3,200 a year, which is a fraction of one junior developer. A team already on ChatGPT Business gets Codex for nothing extra and can run the trial before adding Claude seats.

Where teams overspend is buying Max for everyone because one person hit a limit. Mixed seats exist for that, and the usage view in the admin console shows who actually needs the bigger seat.

An agent that edits files and runs commands is a different risk from an assistant that suggests lines. Three controls, in place before the first agent seat: a rule that agent changes go through the same pull-request review as human ones, with no direct pushes to main; scoped credentials, so the agent cannot reach production secrets from a developer’s laptop; and a written note in the acceptable-use policy on what code and data may be given to which tool, on which plan. None of it is onerous, and all of it is what a client’s security questionnaire will ask about within the year.