Every major coding agent now bills teams by usage. The seat buys an allowance or a pool, and the work beyond it is metered by tokens. That makes seat price a poor guide to cost. To compare tools, price every session at one reference rate, on the same kind of work, and check what each session produced.
As of November 2026. Vendor terms change often, so each row of the table below carries the date we last checked it.
The seat is a down payment
Anthropic's Claude Code cost docs put it in one line: "Claude Code charges by API token consumption." The same page reports that across enterprise deployments the average is around $150 to $250 per developer per month. A Team Standard seat lists at $20 a month on annual billing. The seat price is the smallest number in that comparison.
The others work the same way. Copilot spends AI Credits, consumed by tokens at each model's API rate. Codex prices its models in credits per million tokens. Cursor's plans draw from a usage pool and bill third-party models such as Claude and GPT separately.
On an allowance plan, the invoice shows what spilled past the allowance, not what the team used. Anthropic's docs say it directly for Team and Enterprise: "Usage inside the seat allowance isn't metered in dollars." Copilot's inline completions and next edit suggestions sit outside AI Credits altogether. Compare two tools by invoice and you're mostly comparing whose allowance was bigger.
Four tools, four units of account
Each row was checked against the vendor's own pages on the date shown. Where a page didn't state something, we left it out rather than guess.
| Tool (team plan) | What the seat includes | What's metered past it | Team spend controls | Last verified |
|---|---|---|---|---|
| Claude Code (Team, Enterprise) | Team: Standard seat $20/mo billed annually ($25 monthly); Premium $100 ($125) with 5x the usage. Enterprise: $20/seat billed annually plus usage at API rates, 20-seat minimum. Pricing | Team: a per-seat allowance on five-hour and weekly windows, then optional usage credits. Docs | Spend limits by org, group or member; spend report per user and model (CSV); Analytics API on Enterprise. Docs | Nov 3, 2026 |
| Codex (ChatGPT Business, Enterprise) | Business: $20/user/mo billed annually ($25 monthly), with access to ChatGPT and Codex. Enterprise: custom. Pricing | Credits, priced per model per million tokens; workspaces can buy more. Pricing | Monthly credit limits by seat type and per user. Help center | Nov 3, 2026 |
| Cursor (Teams, Enterprise) | Teams from $30/user/mo, drawing on a weekly usage pool. Existing customers keep their current plan. Pricing, help | Third-party models (Claude, GPT) at API list price from prepaid credit; top-ups when the pool runs out. Help, pricing | Usage analytics; spending caps on third-party usage; group-level spend limits on Enterprise. Pricing | Nov 3, 2026 |
| GitHub Copilot (Business, Enterprise) | 1,900 (Business) or 3,900 (Enterprise) AI Credits per user per month, pooled at the billing-entity level. Seat price: see vendor page. Docs | AI Credits at $0.01 each, from tokens at each model's API rate. Inline completions aren't billed in credits. Docs, blog | Budgets at user, cost-center, org and enterprise level; a policy that blocks usage past the pool. Docs | Nov 3, 2026 |
Compare at one reference price, on the same work
Price lists move, and none of these units convert into the others. The method for comparing them can stay put.
Convert everything back to tokens. Every one of these units is tokens underneath. Price each session's tokens at published per-token rates, the same way for every tool. The result won't match any invoice, and that's fine. It's a common ruler, and it counts the usage the invoice leaves out.
Hold the kind of work constant. Each kind of work costs a different amount in any tool. If one team uses Codex mostly for refactors and another uses Claude Code mostly for debugging, comparing the two averages tells you about the teams. Compare Codex on refactors with Claude Code on refactors.
Split the model out from the tool. Inside one tool, the spread between models is wide. Codex's rate card charges 50 credits per million input tokens for GPT-6.1 Sol and 2.5 for GPT-6 Luna, a 20x gap. A premium model on routine edits makes a cheap tool look expensive. If you skip this step you can end up switching tools when the real fix was a default model.
Divide by what came out. Spend per authored pull request is the simplest check. Next to it, look at the share of spend in sessions that had to redo their own work. A tool that costs more per session but rarely loops can still be the cheaper way to get a change merged. It's the argument from our tokenmaxxing post: tokens are an input, so measure what they bought.
Keep the comparison at the level of teams, tools and kinds of work. A per-engineer cost ranking mostly tells you who drew the hard problems.
Cursor looked 44% cheaper, until the spend was split by kind of work
A fictional 40-engineer company, call it Halvard Freight, runs Claude Code and Cursor side by side for a month before a renewal. The invoices for usage beyond the seats come to $3,200 for Claude Code and $1,800 for Cursor, because most Cursor use stayed inside its pool. On invoices, Cursor looks 44% cheaper.
Priced at one reference rate and split by kind of work, it looks different:
| Kind of work | Tool | Spend at reference prices | Authored PRs | Spend per PR | Share of spend in heavy rework |
|---|---|---|---|---|---|
| Debugging | Claude Code | $6,000 | 120 | $50 | 15% |
| Debugging | Cursor | $4,200 | 70 | $60 | 35% |
| Small features | Claude Code | $4,500 | 90 | $50 | 20% |
| Small features | Cursor | $3,000 | 100 | $30 | 10% |
Add the rows up and Cursor is still cheaper overall: $7,200 for 170 PRs is about $42 a PR, against $10,500 for 210 PRs, or $50, for Claude Code. At reference prices the gap is about 15%, not the 44% the invoices showed, and it reverses on debugging ($60 against $50), where 35% of Cursor's spend went into sessions that redid their own work. Halvard makes Cursor the default for small features and reads a dozen Cursor debugging sessions before renewal. If the rework traces to a missing instruction, they fix that first. If it doesn't, debugging moves to Claude Code.
What changed recently
These are the changes behind the table. They'll date; the method above shouldn't.
- June 1, 2026. Copilot replaced premium requests with GitHub AI Credits. GitHub announced the change on April 27.
- October 7, 2026. Anthropic announced monthly API credits for Max and Team subscribers, rolling out that week. Anthropic's help center says they don't cover interactive Claude Code sessions, and Team credits pool to at most $500 a month.
- October 2026. Cursor's new plans launched inside The X Subscription, bundled with Grok and X Premium (The New Stack, October 13). Third-party models are billed separately at API list price; existing customers can keep their old plans.
- October 14, 2026. GPT-5.5 retired from Codex for users signed in with ChatGPT. API-key usage wasn't affected.
Where VibeCheck fits
VibeCheck gives you the numbers for this comparison across Claude Code, Codex and Cursor. A stop hook and background sync, installed with one command or through your MDM, capture each session on the engineer's laptop, and VibeCheck prices it from its own tokens at published reference prices. Those figures are estimates, not invoice amounts. Cursor's desktop agent reports less than the other two, and VibeCheck labels those sessions as partial. It doesn't capture Copilot, or agents that run in a vendor's cloud rather than on the laptop.
On Pro, Usage groups spend by coding tool, model, team, person and skill, and shows the token mix by kind of work. Filter Sessions by kind of work and tool to see what the matched sessions cost, so you can compare Claude Code, Codex and Cursor on your own work and check whether the model fits the work. Connect the GitHub App for cost per authored PR. Sessions shows what heavy-rework sessions cost and where engineers corrected the agent. A provider console shows one vendor's usage; here's how that compares.
The free plan covers up to five engineers, with each session's cost and spend by model. Team views and cost per PR come with Pro, at $30 per engineer per month billed annually. See pricing, or get started for free.