Cursor vs Copilot vs Claude Code: the real cost per engineer

Here is the short answer. GitHub Copilot Business is $19/user/month and Enterprise is $39/user — but Enterprise requires GitHub Enterprise Cloud at about $21/user, which makes the real seat roughly $60. A seat is a paid license for one person. Cursor Business is $40/user, with negotiated rates seen at $30–45, and $25–35 at volume. Claude Code Team Premium is $100/seat with a five-seat minimum, and spend seen in practice runs $150–250 per developer per month. All prices read in July 2026 — check them before you build anything on them.

This post does not tell you which tool is better. Spendassay takes no view on that, and a cost auditor with a favorite vendor is not an auditor. What it does tell you is how the three tools charge. That decides whether your bill tracks your seat count or your workload, and those are two budget problems, not one.

Three tools, three ways of charging

The list prices are not comparable, because the three products charge for use in different ways. Sort them by how they charge and the comparison starts to work.

Mostly fixed seat. Copilot Business at $19/user behaves closest to normal software. You buy seats, you pay for seats, and the bill lands within a few percent. There is an on-demand part — overage runs $0.01 per credit, per GitHub's plans page — but for autocomplete-heavy use, the seat is most of the number.

Seat plus included usage that falls off a cliff. Cursor Business at $40/user bundles some usage into the seat, then switches to on-demand billing once that allowance is used up. The point here is not the price, it is the shape. The bill is flat, flat, flat, then it is not. Teams that pick up agent workflows mid-quarter find the cliff after they have gone over it.

Seat plus usage with no real cap. Claude Code Team Premium at $100/seat sets a floor, not a ceiling. Reported average use of roughly $13 per developer per active day means a heavy user can clear the seat price in a week. This is the structure where the sticker tells you least.

Same category, same buyer, three links between what your engineers do and what you pay. That is the real finding.

Three tools, three meters — the cheapest seat and the cheapest bill are rarely the same vendor.

The list prices, dated

Checked July 2026. Each one of these will move.

| Tool | Plan | List price | Usage charge on top | |---|---|---|---| | GitHub Copilot | Business | $19/user/mo | On-demand credits, $0.01/credit | | GitHub Copilot | Enterprise | $39/user/mo + ~$21 GHEC = ~$60 effective | On-demand credits | | Cursor | Business | $40/user/mo | Included usage, then on-demand | | Cursor | Pro / Pro+ | $20 / $60 per mo | Included usage, then on-demand | | Claude Code | Team Premium | $100/seat/mo, 5-seat minimum | Usage-driven; observed $150–250/dev/mo | | Claude Code | Max 5x / Max 20x | $100 / $200 per mo | Usage-driven |

The GitHub Enterprise Cloud line matters more than its size suggests. Most comparisons list Copilot Enterprise at $39 and stop, which puts it below Cursor Business at $40. Add the platform license it requires and Copilot Enterprise lands around $60. That is above Cursor list, and above most negotiated Cursor rates. The ranking flips on a line item that is not on the comparison page.

For negotiated pricing, Vendr's Cursor marketplace data is the most useful public reference. It shows a median yearly contract value of $111,408 across 76 documented purchases, ranging from $19,200 to $395,040. Discounts of 15–25% were seen against a yearly commitment, and 25–40% when a few levers were combined. That is real transaction data, not a vendor's own claim, and it tells you the list price is a starting position.

A worked example: 100 engineers

Assume 100 engineers, for a full year, with one tool across the whole company so the structures show. Seat math here is counted — read straight off published prices, with no guessing. Meter math is estimated — worked out from assumptions you can see and change, which you should replace with your own data.

Profile A — light inline use. Autocomplete-heavy, some chat, few or no long-running agent tasks.

| Tool | Per engineer/mo | 100 engineers/yr | |---|---|---| | Copilot Business | ~$19–25 | ~$23,000–30,000 | | Cursor Business (negotiated $30) | ~$30–35 | ~$36,000–42,000 | | Claude Code Team Premium | ~$100+ | ~$120,000+ |

The order is clean and matches the sticker: Copilot, then Cursor, then Claude Code. At this level of use, list price predicts the bill well, and the premium seats are paying for capacity nobody is using.

Profile B — heavy agent use. Long-running agents, large context windows, multi-file refactors, frequent tool calls.

| Tool | Per engineer/mo | What drives it | |---|---|---| | Copilot Business | $19 seat + on-demand credits | Seat stops predicting the bill once the allowance runs out | | Cursor Business | $40 seat + on-demand | Same cliff, different allowance | | Claude Code Team Premium | $150–250 observed | Usage-led from the start |

At this level the seat prices fade into noise. DX's June 2026 pricing analysis puts total cost per engineer — seat plus tokens — at $200–600/month for teams mixing inline and agent tools. A token is the unit AI vendors bill for, roughly a few characters of text. That band is wide enough to swallow the whole 5x sticker gap between Copilot Business and Claude Code.

The honest version of Profile B is that no table of numbers holds up here. The on-demand part depends on your token volume and your model mix, and no published dataset settles that per tool. What survives is the shape. At light use the sticker ranking holds. At heavy use the meter dominates, and the ranking turns into a question about your own consumption.

Why the population data is so wide, and what that tells you

Even the most credible public figure is a wide band, not a point. DX puts total cost per engineer at $200–600 a month — seats plus tokens, on teams mixing inline and agent tools. That is a 3x spread inside a single source.

The width is the finding. The bottom of that band is companies still mostly on inline tools; the top is teams that have moved to agents and put the tail on the meter. A sticker comparison hides all of it. And every dollar of that band was spent by some other company, which is why it sets a range to check yourself against and answers nothing on its own.

So the only comparison that answers your question is all-in cost per engineer per month, per tool, on your own usage. That is an arithmetic problem, not a research problem. But it needs the token data sitting next to the seat data, and that is the part most companies do not have. How each vendor gives up that data varies a lot, and what each vendor's admin console actually shows covers the gaps.

What this means for a buying decision

Three results, none of which involve ranking the tools.

Budget the structure, not the sticker. Settle on a mostly-fixed-seat product and your AI line behaves like normal software, so it belongs in a buying process. Settle on a usage-led product and it behaves like cloud spend, so it needs a run-rate owner and a monthly review.

Price the whole stack, not the product page. The GitHub Enterprise Cloud line is the obvious example. Check for platform licenses you are required to hold, the smallest seat count you can buy (Claude Code Team Premium's five-seat floor matters at small scale), and what happens to the included allowance when you change tiers.

Assume you will run more than one. Most companies end up with an inline tool and an agent tool rather than one winner. That raises a separate and more expensive question about whether the overlap is worth paying for. The comparison above is single-tool for clarity. Your bill probably is not.

The honest caveats

This is a snapshot, and a fragile one. Each price here was read in July 2026 from vendor pricing pages, and AI tool plans get restructured faster than almost any software category. Allowances get re-cut, credits get redefined, tiers get renamed, and a plan that was cheapest in July can be mid-pack by October. Check before you commit. The negotiated Cursor rates are observed ranges from third-party transaction data, not a rate card you are entitled to. The Profile B table gives shape rather than numbers, because no public dataset settles on-demand use per tool per usage profile. Treat it as a structure to fill in with your own billing data, not a result. And nothing here judges output quality, developer preference, or fit. How a tool charges is one input to a decision that has several.

Find your recoverable AI spend

The comparison that decides this is your own all-in cost per engineer per tool — seats plus tokens, on your own usage. Spendassay reads both sides from Cursor, GitHub and Anthropic and returns the per-engineer number, so the ranking comes from your bill rather than a pricing page.

`Start free` → /login?src=blog_cursor-copilot-claude-code-real-cost