AI Coding Assistant Pricing Compared: Per-Seat vs Per-Token vs Hybrid (2026)
If you are deciding where AI genuinely helps across the software lifecycle before you commit budget, start with our piece on AI in the SDLC. Here we focus on one practical question: what these tools cost, and why the bill is so hard to predict.
The three pricing models (this is the part that matters)
The sticker price tells you almost nothing until you know which model a tool bills on. There are three, and they behave completely differently as a team scales up its usage.
- Per-seat. A flat monthly fee per developer, used or not. Predictable and easy to budget, but you pay for the seat whether the developer runs one completion a day or a thousand. Historically how GitHub Copilot and Gemini Code Assist worked.
- Per-token / usage. You pay for what the agent actually consumes - tokens, credits, or agent runs. Scales with real use, which means it can stay tiny on a light week and spike hard during a heavy agentic sprint. Claude Code and anything driven through a raw model API sit here.
- Hybrid. A seat fee that includes an allowance, then metered overage once you cross it. The model teams understand least and get surprised by most, because the bill looks flat until the month it isn’t. Cursor and AWS Kiro work this way, and - the important 2026 shift - GitHub Copilot moved here too: since 1 June 2026 a Business seat is still $19 but now includes $19 of AI credits, with overage billed at each model’s API rate.
The headline for 2026: the industry has converged on metered usage on top of a seat. A flat per-seat world is gone. That is why “how much is Copilot” no longer has a single answer, and why the FinOps section at the end matters more than the price table.
GitHub Copilot vs Microsoft Copilot: untangling the name
Before any price comparison, clear up the single most confusing thing in this market - because Microsoft put the word “Copilot” on at least eight different products, and two of them get mixed up constantly.
- GitHub Copilot is the developer coding tool - completions, chat, and agents inside your IDE. This is the one that belongs in a comparison with Cursor and Claude Code. Pricing (2026): a free tier, Pro at $10/user/month, Business at $19/user/month, and Enterprise at $39/user/month.
- Microsoft 365 Copilot is a different product entirely - the AI assistant inside Word, Excel, Outlook, and Teams, aimed at office workers, not developers. It costs $30/user/month on an annual commitment, on top of a qualifying Microsoft 365 licence you already pay for. It has nothing to do with writing code.
So “we’re rolling out Copilot” from your IT department and “we’re trialling Copilot” from your engineering lead can mean two unrelated tools with different buyers, different budgets, and different bills. And it does not stop there: Microsoft also ships Copilot Pro (a $20/month consumer add-on), Copilot Studio (consumption-based agent builder), Copilot for Security, and more - the portfolio spans eight-plus separately metered SKUs. When someone says “Copilot,” the first question is always which one.
| GitHub Copilot | Microsoft 365 Copilot | |
|---|---|---|
| What it is | AI coding assistant in the IDE | AI assistant in Word/Excel/Teams |
| Who it is for | Developers | Office/knowledge workers |
| Entry price | Free / $10 Pro / $19 Business | $30/user/mo (annual) |
| Prerequisite | None (a GitHub account) | A qualifying M365 licence |
| Belongs in this guide? | Yes | No - different product |
For the rest of this comparison, “Copilot” means GitHub Copilot, the coding tool.
Trying to work out what AI coding tools will actually cost your team?
Book a free 30-min call
Tool by tool (as of November 2026)
A tour of what each coding assistant charges and, more importantly, what drives the bill. Prices move fast in this market - treat these as the shape of each model and confirm the current number on the vendor’s page before you budget.
GitHub Copilot - per-seat, now with metered overage
Free tier, Pro at $10, Business at $19, Enterprise at $39 per user per month. Since June 2026 the Business seat includes $19 of AI credits and bills overage at model API rates, so heavy agent use pushes past the flat seat. Best fit: teams standardised on GitHub who want the shortest path and a predictable floor.
Cursor - hybrid, the most variable
Individual plans from $20 up to $200/month, with team seats that bundle an allowance of requests then charge for overage. Cursor is the most variable of the group: light use stays near the seat price, heavy multi-file agent work can climb quickly. Best fit: teams that want raw in-editor speed and will watch the usage.
Claude Code - consumption, the most predictable ceiling
Included in Claude Pro ($20/month), Max ($100 and $200), and Team seats (from $25/seat, billed annually), with Enterprise at a seat fee plus API usage. It is consumption-driven, but in practice it has the most predictable ceiling of the usage-based tools. Best fit: terminal-driven agentic work - codebase navigation, test generation, log triage.
OpenAI Codex - bundled into ChatGPT tiers
Codex rides along with ChatGPT subscriptions: Plus ($20), Pro ($100 and up), and Business/Enterprise, plus raw API access if you wire it into custom agents. Calling the API directly for custom agent workflows is where costs scale fastest. Best fit: teams already living in the ChatGPT/OpenAI ecosystem.
Gemini Code Assist - per-seat, Google-native
Per-seat Standard and Enterprise tiers, tightly integrated with Google Cloud. Best fit: teams on Google Cloud who want the IDE assistant to line up with the rest of their stack. (If that is you, the Vertex AI / Gemini platform picture is the wider context.)
AWS Kiro - credit-based, spec-driven
Kiro uses a credit model with a free tier and paid Pro/Power/team plans, billing by credits tied to agent interactions rather than a flat seat alone. It is a different animal - spec-driven rather than prompt-first - so we cover it in depth separately in what AWS Kiro is and where it fits. Best fit: teams on AWS who want AI-written changes to be reviewable as specs.
The honest comparison table
The thing a vendor pricing page will not tell you in one view: what model each tool bills on, and what makes the bill move.
| Tool | Pricing model | Entry (per user/mo) | What scales the bill |
|---|---|---|---|
| GitHub Copilot | Per-seat + metered overage | Free / $10 / $19 / $39 | Agent/credit usage above the seat allowance |
| Cursor | Hybrid (seat + overage) | $20 to $200 | Requests and agent runs above the included allowance |
| Claude Code | Consumption | $20 (Pro) / $25+ (Team) | Tokens consumed by agent sessions |
| OpenAI Codex | Bundled in ChatGPT + API | $20 (Plus) and up | API calls for custom agent workflows |
| Gemini Code Assist | Per-seat | Standard / Enterprise seat | Mostly flat; tier and add-ons |
| AWS Kiro | Credit-based | Free / Pro / Power | Credits tied to spec and agent interactions |
Token economics for a software team
Here is where the model choice shows up on the invoice. Take a 20-developer team:
- On a light month (completions and occasional chat), a per-seat tool like Copilot Business is almost exactly 20 x $19. The usage-based tools stay well under their ceilings. Per-seat looks expensive because you pay for idle seats; usage-based looks cheap.
- On a heavy month (agents running multi-file refactors, test generation, and log triage all day), the picture flips. The usage-based and hybrid tools climb - vendor-reported heavy usage lands around $100 to $250 per developer per month regardless of which tool you pick - while the flat portion of a per-seat plan holds steady and only the overage moves.
The practical takeaway: at low, steady usage, per-seat is predictable and often cheaper; at high agentic usage, every model converges on roughly the same real cost, and the risk shifts from “how much per seat” to “how do we stop a runaway agent month.” That is a FinOps problem, not a procurement one.
Controlling the spend (the part most comparisons skip)
Picking the cheapest seat is the easy 10%. The other 90% is keeping the bill predictable once agents are doing real work across a team. The same cost-governance discipline we apply to cloud infrastructure applies here:
- Meter it. You cannot manage what you cannot see. Capture usage per developer, per team, and per repository, not just the one line on the vendor invoice.
- Attribute it. Tie spend to teams and projects so a runaway month has an owner and a cause, the same way you would tag cloud resources.
- Budget and alert. Set a per-team ceiling and alert before it is breached, rather than discovering the overage on next month’s bill.
- Standardise the tool and model choice. Left to individual preference, a team ends up paying for three overlapping tools and burning premium-model tokens on tasks a cheaper model would handle. Decide the default once.
- Watch for sprawl. One developer with an agent is a productivity story; a whole org running agents, subagents, and background tasks with no shared policy is the agent sprawl pattern - now showing up on the AI-tooling bill instead of the infrastructure one.
This is cloud FinOps applied to AI tooling: the tools changed, the discipline did not. If cost control across your whole platform is the real question, that is our FinOps work.
Summary
Three things to take away:
- The model matters more than the sticker price. Per-seat is predictable and cheaper at light, steady use; per-token and hybrid scale with real usage and can spike. In 2026 almost everyone meters above a seat allowance, so the seat price is a floor.
- “Copilot” is two different products. GitHub Copilot is the coding tool ($19 Business); Microsoft 365 Copilot is the Office assistant ($30, on top of an M365 licence). Know which one is on the table before you compare anything.
- At team scale, the bill is a FinOps problem. Once agents work all day, the tools converge on roughly the same real cost and the risk moves to runaway usage - so metering, attribution, and budgets matter more than the entry price.
Not sure what AI coding tools will cost your team?
Book a free 30-minute call. We help teams pick the right AI coding stack for how they actually work, model the real per-seat and per-token costs, and put the metering and budgets in place so the bill stays predictable.
