A Copilot seat is a floor price, not an estimate. At list on 9 September 2026, the four seats a large enterprise actually shortlists are $19 for GitHub Copilot Business, $20 for ChatGPT Business on annual billing, $30 for Microsoft 365 Copilot on an annual commitment, and not published for Gemini Enterprise. Every one of them meters something the seat does not cover.
In the one case where you can compute the entire bill from published rates — GitHub — a developer who runs one agent session per working day costs $174 a month against a $19 seat. Not because anyone is gouging: because completions are free forever and agents are billed on tokens, and the seat price was never a prediction of which one your engineers would use.
Here is the recommendation, before the analysis. Buy assistant seats on measured weekly actives, buy coding seats on task mix, and put a hard spending cap on the agent surface before your first uncapped invoice. Treat the licence line as well under half the twelve-month number, not the whole of it.
| What you are signing for | Microsoft 365 Copilot | GitHub Copilot Business | Gemini Enterprise Standard | ChatGPT Business |
|---|---|---|---|---|
| Published seat price | $30/user/mo, paid yearly ($31.50 monthly on annual commitment) | $19/user/mo | Not published — contact sales | Standard $20/user/mo annual, $25 monthly; Premium $100/$125 |
| Prerequisite | A qualifying Microsoft 365 plan | A GitHub organisation | A Google Cloud project | None |
| Included consumption | Agent usage "no charge" for licensed users, under fair-use limits | 1,900 AI credits (= $19) | $10 credit per user per month | Per-seat feature limits |
| How the pool behaves | Tenant-wide prepaid capacity packs | Org-wide pool, forfeited at month end | Org-wide, enforced on a rolling 7 days, forfeited weekly | Workspace shared credit pool |
| Overage rate | $0.01/credit pay-as-you-go; $0.008 prepaid | $0.01/credit at published per-model token rates — cached input bills at 10% of input | Agent Platform rates | Purchased credits |
| Storage meter | Microsoft 365 Archive at $0.05/GB/mo above licensed quota | None | $5 per GiB per month above 30 GiB per licensed user, pooled | Not published |
| What happens at the cap | Custom agents disabled at 125% of prepaid capacity | Blocked, or metered — admin's choice | Draws stop until the week rolls | Feature limits bind |
| Never metered | — | Code completions and next edit suggestions | — | — |
The Four Bills You Actually Sign
A Copilot rollout generates four separate charges, and most business cases model one. The licence is the only one that scales with headcount. The other three scale with behaviour, corpus size and legal exposure — none of which appear in a seat count.
One: the licence. Microsoft 365 Copilot is $30.00 per user per month paid yearly, or $31.50 paid monthly on an annual commitment, and it is an add-on — a separate qualifying Microsoft 365 or Office 365 plan is required before you can assign it. GitHub Copilot Business is $19 per granted seat per month, Enterprise $39.
Two: consumption. Every vendor has now separated "access" from "work done." GitHub bills Copilot Chat, Copilot CLI, the Copilot cloud agent, Copilot Spaces, Spark and third-party coding agents against a credit allowance where one credit equals one US cent. Microsoft bills agent activity in Copilot Credits at $0.01 per message pay-as-you-go, or $200 per pack per month for 25,000 credits prepaid — effectively $0.008 a credit. Google gives each Gemini Enterprise Standard seat a $10 credit per user per month and charges Agent Platform rates beyond it.
Three: storage. An assistant that grounds in your corpus has to index your corpus, and indexes are billed by the gigabyte. This is the line that appears in no seat model, because it does not scale with people.
Four: compliance. Every prompt and every response becomes a retained, discoverable, exportable record. That is not a new SKU. It is new volume on a compliance estate you already pay for, and its growth is the exact number your business case is trying to increase.
GitHub Copilot: The Cheapest Seat and the Steepest Meter
GitHub Copilot is the right buy for developers and the hardest bill to forecast, and both facts have the same cause. On 1 June 2026, GitHub replaced premium request units with GitHub AI Credits, holding seat prices at $19 for Business and $39 for Enterprise and setting each seat's monthly allowance to exactly its price — 1,900 and 3,900 credits respectively.
Code completions and next edit suggestions are not billed in AI credits and remain unlimited on all paid plans. That is genuinely generous, and it is why a developer who uses Copilot the way developers used it in 2024 costs exactly $19 a month, forever. Agents are the other story.
Credits are consumed on input, output and cached tokens, and the rate card prices each of those separately. Checked today: GPT-5.4, the default, is $2.50 per million input tokens, $0.25 per million cached input and $15.00 per million output. Claude Opus 5 is $5.00, $0.50 and $25.00, plus $6.25 per million for cache writes. Claude Sonnet 5 is $2.00, $0.20 and $10.00, plus $2.50 for writes.
That cached column is the one to model, and almost every forecast ignores it. It is a tenth of the input rate, and an agent loop lives inside it: every turn re-sends the accumulated context, so cache reads come to dominate the input count. Published teardowns of Claude Code sessions — a different product on the same billing shape — put cache reads at 97% or more of all tokens consumed. Price a repository-scale context at the uncached rate and you will overstate the bill roughly twofold.
Now price one agent session that reads a repository, plans, edits and iterates — call it four million input tokens and 400,000 output tokens, which the coding agents reach without effort. Assume a deliberately conservative 90% of that input is cache reads, well below what the teardowns report:
- On GPT-5.4: $1.00 uncached input + $0.90 cached + $6.00 output = $7.90, or 790 credits.
- On Claude Opus 5, where cache writes cost more per token than fresh input: $2.50 + $1.80 + $10.00 = $14.30, or 1,430 credits.
Note what dominates. Output is 76% of the GPT-5.4 session and 70% of the Opus one, against a context ten times its size. Forecast output tokens, not context size — the intuition that a big repository is the expensive part is precisely backwards, and it is the intuition every seat-versus-consumption discussion starts from.
A Business seat's entire month is 1,900 credits. One Opus 5 agent session is three-quarters of it. Bill the same two sessions with no cache reuse at all and they come to $16.00 and $30.00 — treat that as the ceiling, not the estimate.
Take the 400-developer fleet
Four hundred Business seats cost $91,200 a year. Credits pool at the organisation level rather than per user — 400 seats means one shared pool of 760,000 credits, worth $7,600 a month — and unused credits do not carry forward; the pool resets at UTC midnight on the first of each month and the remainder is forfeited.
Assume a quarter of that fleet — 100 developers — runs one 790-credit GPT-5.4 agent session per working day. That is 1.74 million credits a month, or $17,380 of consumption against $7,600 of pooled allowance. The metered overage is $9,780 a month, $117,360 a year, on top of the $91,200 licence.
Total: $208,560 a year, 2.3x the licence line. Per heavy developer, $174 a month against a $19 seat — a multiple of nine. Model the same fleet at uncached input rates and it comes out at $422,400 and a multiple of 18.5. That gap between two defensible-looking spreadsheets is the largest single modelling error available to you here, and it runs toward over-provisioning.
The pooling is the trap inside the trap. Three hundred completion-only developers donate their $19 each to the hundred who are burning agents, so the average seat looks like it costs $43 a month and nothing in the org-level view identifies the tail. You cannot manage a distribution you have averaged away.
The 1 September repricing nobody's budget shows
Existing Business and Enterprise customers received elevated allowances of 3,000 and 7,000 credits for June, July and August 2026. Those expired on 1 September. Business fell from 3,000 to 1,900 and Enterprise from 7,000 to 3,900 — cuts of 37% and 44% with no change in seat price.
On the 400-seat fleet above, that alone removed $4,400 a month — $52,800 a year — of included allowance. Anyone who sized a 2026 budget during the promotional window sized it against a number that no longer exists, and the first honest invoice lands in early October, straight into planning season.
Developers noticed before finance did. In GitHub's own community forum, one reported using 100% of a Pro+ allowance "in just 2 days of normal work"; another that a four-hour pairing session wiped out two weeks of credits; a third that an organisation's bill "went from $295 in june to $1K+ in July." Those are anecdotes, not a benchmark. They are the right shape, and the shape is what your model has to survive.
The honest steel-man: GitHub's rate card is fully published, per model, per token type, and admins get user-level, cost-centre, organisation-level and enterprise spending caps with the most restrictive one binding first. Everything above is knowable in advance. It just requires you to forecast token consumption, which almost nobody who signs the contract can do.
Microsoft 365 Copilot: Agents Are Free Until They Are Not
Microsoft 365 Copilot is the most predictable of the four, and its waste is idle seats rather than runaway consumption. For a licensed user, the entire Copilot Studio billing table reads "No charge": classic answers (1 credit), generative answers (2), agent actions (5), tenant graph grounding (10), agent flow actions (13 per 100), content processing (8 per page). Building and adding agents is included in the licence.
That inclusion carries four exclusions, and they are where the bill comes from.
Unlicensed users pay full freight. Microsoft's own worked example: a tenant-graph-grounded sales agent averaging four generative answers and four grounded calls per run — 48 credits — serving 50 licensed and 100 unlicensed users. Microsoft bills only the 100, at 4,800 credits a day. That is $48 a day pay-as-you-go, $17,520 a year for one agent, or about $14,000 on prepaid packs. Their external customer-support example runs 7,200 credits a day for 900 website visitors: roughly $21,000 a year at pack rates. Any agent that touches a customer, a contractor or a frontline worker without a Copilot licence is a consumption product wearing a subscription's clothes.
Computer-using agents are excluded outright. The footnote is one line — computer-using agent usage "is not included in the Microsoft 365 Copilot USL" — and it bills at the 5-credit agent-action rate. If your 2027 plan involves agents that drive applications rather than answer questions, that plan is outside your licence.
The trigger decides, not the agent. Agent flows are free for licensed users only on the "When an agent calls the flow" trigger. Every other trigger — scheduled, event-driven, autonomous — bills at 13 credits per 100 actions. The same flow, doing the same work, is free or metered depending on what woke it up.
Reasoning models bill twice. Microsoft charges the feature rate plus premium text and generative AI tools at 10 credits per 1,000 tokens. A generative answer from a reasoning model over 4,000 tokens is 2 + 40 = 42 credits, or $0.42 a question. Turn on deep reasoning across a department and you have changed the unit economics of the whole deployment.
And there is a hard stop. Enforcement triggers at 125% of prepaid capacity and disables custom agents. Users see "There is a billing issue" or "This agent is currently unavailable. It has reached its usage limit." A production agent failing closed on a billing threshold is an availability incident with a finance root cause — the same class of failure as an agent retry loop riding into a provider outage.
Note also that "fair usage limits" governs the free inclusion, and Microsoft "reserves the right to update the limits as the product evolves." The most valuable term in your Copilot contract is the one with no number attached to it.
Gemini Enterprise: A Monthly Credit Enforced Weekly
Gemini Enterprise has the most under-modelled meter of the four, and the mechanic is stated plainly in Google's own documentation. Quotas "are expressed as a monthly per-user amount" but "are enforced on a rolling seven-day basis as a shared pool", where the weekly pool is the per-user monthly value divided by four, multiplied by purchased seats. Unused quota "doesn't roll over to the next week."
For 5,000 Standard seats at a $10 monthly credit, that is $12,500 per rolling week — not $50,000 a month. The distinction only matters if your usage is uneven, which is to say: it matters.
Work a quarter-end. Weeks one to three run at $5,000 of consumption each; week four, with the close and the board pack, runs at $25,000. Total usage for the month is $40,000 against a notional $50,000 allowance — and you still pay. Weeks one to three forfeit $22,500 of unused quota, and week four generates $12,500 of overage at Agent Platform rates. You underspent your monthly allowance by 20% and received an overage invoice. No seat-count model produces that result, because no seat-count model knows the pool resets weekly.
Then storage. Google's editions table gives Business 25 GiB, Standard 30 GiB, Plus 75 GiB and Frontline 2 GiB — per licensed user, pooled across the project — with storage and data indexing overage at $5 per GiB per month. Google's own worked example: 100 Standard licences plus 50 Plus licences is one combined pool of 6,750 GiB.
So this meter is a ratio, not a total — it bills when your indexed corpus exceeds roughly 30 GiB per Standard seat. On 5,000 seats that is 150,000 GiB of headroom, and a one-terabyte corpus costs nothing. On 200 seats it is 6,000 GiB, and the same 10 TiB estate that a large deployment absorbs for free sits 4,240 GiB over: $21,200 a month. The exposed buyer is the small-seat, large-corpus one — a legal, compliance or research group indexing everything for a few hundred people.
Two fairness points. First, $5 per GiB per month buys a live, queryable search index, and it is not comparable to cold archive storage; do not put it next to Microsoft 365 Archive's $0.05 per GB per month and call it a hundredfold difference, because they are different products. Second, Google bundles Gemini into Workspace Business plans at roughly $7 to $22 per user per month with no AI add-on at all — which for a lot of organisations is the actually correct purchase, and the one that never appears in a Gemini Enterprise proposal. Google's agent platform meters the rest, and its metered session memory is a separate line again.
ChatGPT Enterprise: The One You Cannot Model
ChatGPT Business is the cheapest credible assistant seat and the easiest to buy; ChatGPT Enterprise is the one you cannot put in a spreadsheet before signature. Business lists at $20 per seat per month on annual billing and $25 monthly, with a two-seat minimum, and a contractual commitment that OpenAI will not train on your data. Enterprise is quote-based — everything in Business plus SCIM, data residency, expanded context, analytics and volume terms, priced by sales and billed annually.
Advanced capabilities — Deep Research, reasoning models, image generation, advanced voice, Codex — run on per-seat limits that draw from a purchased workspace credit pool once a user exceeds them. The structure is the same as everyone else's. What is missing is a published rate card you can multiply by a forecast.
Since 25 August 2026 there is a second answer to hitting those limits, and it is the strongest confirmation of this article's thesis in any of the four products: Premium seats at $100 per seat per month billed annually, $125 monthly — five times the Standard usage allowance, the five-hour limit removed, and mixable with Standard seats inside a single workspace. OpenAI's answer to "the meter binds" is a seat costing five times as much. So $20 is an entry tier, not a price, and a Business business case that multiplies headcount by $20 has already assumed away the decision. Budget the mix.
That is a real cost, not a rhetorical one. Every other product in this comparison lets a finance owner build a defensible twelve-month number from vendor documentation alone. For Enterprise you get a quote, and a quote is a negotiation position, not a model. Procurement reports through 2026 converge on a range rather than a price, which is exactly what you would expect from an opaque list.
Buy Business if you want assistant seats without a Microsoft dependency — at $20 for Standard, with a share of the fleet budgeted at $100 for Premium. Take Enterprise to sales only if you already know your usage shape well enough to negotiate against it — and if you do not, run six months on Business first and measure. The same discipline applies to enterprise search platforms, where the permission model matters more than the seat price.
Seats Bought Versus Seats Actually Used
The largest single line in an assistant rollout is the one you can fix without talking to a vendor: seats assigned to people who do not use them. And the reason it survives renewal after renewal is that the vendor's "active user" number is real, published and much weaker than the word implies.
Microsoft's usage report defines Active Users as enabled users "who tried a user-initiated Microsoft Copilot feature, in one or more apps in Microsoft 365 over the selected timeframe," and the selectable timeframes are 7, 28, 90 or 180 days. The FAQ is precise about the bar: opening the Copilot pane in Word does not count, but submitting one prompt does. On the 180-day view, a single prompt in six months makes a user active.
That definition is not dishonest — it is the standard SaaS convention, and the report offers a 7-day view for anyone who wants it. The error is the buyer's: signing a headcount-shaped renewal off a 180-day rate.
Microsoft reported more than 30 million paid Microsoft 365 Copilot seats in its FY26 Q4 results on 29 July 2026, and on the call said the time from deployment to "monthly active usage about 80% across a customers' user base" has fallen from months to days. Monthly and weekly are different questions, and one prompt in 28 days clears the monthly bar. Ask for the 7-day rate, per department.
There is a second trap for anyone building their own dashboard. Microsoft states explicitly that Purview audit log data "isn't intended to be used as the basis for Copilot usage reporting," and that aggregate metrics built on it "might not be consistent" with the official reports. If your FinOps team has wired utilisation to the audit log because that is where the API was, the number will not tie to the one your account team quotes, and you will lose the argument at renewal.
M&T Bank is the useful shape of a real rollout: AI copilots to more than 15,000 employees out of roughly 22,000, after a pilot of about 800, against technology spending above $1.2 billion in 2025 — and the measured result quoted is about six minutes saved per call in the contact centre. That is what a defensible number looks like: a specific task, a specific saving, tied to a population you can count. We have written before about why three different Copilot counts at the same bank were none of them weekly actives.
The Compliance Line Nobody Put in the Business Case
Copilot output is discoverable enterprise data from the moment it is generated, and its volume is the thing your adoption programme exists to increase. Microsoft Purview supports auditing, data classification, sensitivity labels, DLP, insider risk management, communication compliance, eDiscovery, data lifecycle management and Compliance Manager across Copilot interactions. Prompts and responses are stored in the user's mailbox, and eDiscovery reaches them by searching the IPM.SkypeTeams.Message.Copilot.* item class.
That is a good security posture and a real cost. Three things move:
- Retention volume. Every prompt and response is a new mailbox item subject to your retention policies and holds. A 5,000-seat deployment at ten interactions per active user per day is tens of millions of new retained records a year.
- eDiscovery workload. Review sets and exports are priced in analyst hours and platform work. Litigation hold scope now includes a channel that did not exist in the last matter, and the volume in it grows with adoption.
- Storage headroom. Once archived plus active SharePoint storage exceeds the tenant's licensed quota, Microsoft 365 Archive bills $0.05 per GB per month — cheap per unit, and charged only above quota, which is precisely why it never appears until the quarter it does.
None of this argues against buying. It argues for putting your records and legal leads in the business case review before signature, not in the incident review afterwards. The same discipline that applies to per-span observability pricing on agent traffic applies here: retention is a multiplier on a volume you are deliberately growing.
Who Should Not Buy Each One
This is the section a vendor deck will never write, so it is the one worth reading twice.
Do not buy Microsoft 365 Copilot if a large share of your workforce is frontline or deskless. You pay $30 for every assigned seat regardless of use, and the product's value comes from grounding in a document corpus those roles do not produce. Also skip it if your compliance function cannot absorb a step-change in discoverable mailbox items this fiscal year — that is a sequencing problem, not a reason never to buy.
Do not buy GitHub Copilot Business if you need a fixed monthly line and your engineers are agent-heavy. The seat is cheap precisely because the expensive part is metered, and the org-wide pool means the average will look healthy while the tail eats the budget. If you need per-developer chargeback, you need per-user budgets configured on day one, not discovered in month three.
Do not buy Gemini Enterprise if your usage is lumpy, or if your corpus is large relative to your seat count. The weekly-forfeit pool punishes exactly the month-end and quarter-end shape that enterprise work has, and the 30 GiB-per-seat index quota punishes exactly the small team that wants to index everything. Also skip it if your procurement process requires a published price to run a comparison, because there isn't one.
Do not buy ChatGPT Enterprise if you cannot model a unit cost before signature. It is quote-only, and the credit mechanics for Deep Research, reasoning models and Codex are not published in a form you can defend to a CFO. Business at $20 annual is available immediately and costs nothing to leave — just price the Premium seats your heavy users will land on at $100.
And the loser. Judged strictly as a purchase a finance owner has to defend twelve months out, it is Gemini Enterprise for any small team with a large corpus. It is the only one of the four with no published seat price, the only one that forfeits your allowance weekly rather than monthly, and the only one whose index bill is set by a ratio — corpus to seats — that no headcount model contains. Google may well quote you a good price. You cannot check it against anything.
Second place in that ranking goes to GitHub — not for the product, which remains the right buy for developers, but for the transition. Cutting the included allowance 37% on 1 September while holding the seat price is a price rise, and it will not appear in anyone's variance report as one.
Three Moves Before Your Next True-Up
This week. Pull the 7-day active-user rate per department from the Copilot usage report, not the 28-day or 180-day view, and not from the audit log. Set an organisation-level spending cap on GitHub Copilot and a monthly consumption limit on every Copilot Studio agent in the Power Platform admin centre — the cap has to exist before the invoice, because after the invoice it is a credit note negotiation. Then re-run your GitHub forecast on 1,900 credits per Business seat rather than the 3,000 that expired on 1 September.
This month. Classify every agent you run by whether it is invoked by a licensed user, an unlicensed one, or a schedule. That single taxonomy predicts the Microsoft bill more accurately than any usage estimate, because it maps directly to the "No charge" column. Do the same on the coding side: split developers into completion-only and agent-running, and price the second group at the token rates for the model they actually select, not the default — with the cached-input rate applied to the share of context the agent re-sends, and output tokens as the line that actually moves. This is the same exercise as finding the break-even point in a per-phase router and the same discipline as capping cache reads per pull request.
Before renewal. Take three numbers to the negotiation: measured weekly actives, twelve months of consumption at published rates, and your indexed corpus in gibibytes. Ask for consumption to be committed at a discount rather than bought at list, ask for the fair-use limit governing "included" agent usage to be written down with a number, and ask what happens to your included allowance if the vendor changes it mid-term — because one of them just did. Reclaim idle seats before the true-up locks them in, and take the contract-shaped lessons on consumption caps into the room with you.
The Bottom Line
Per-seat software pricing worked because a seat and a user were the same unit: one person, one licence, one predictable cost. Generative assistants broke that identity. A seat now grants access to a meter, and two people holding identical seats can differ by a factor of nine in what they cost — not because one is more valuable, but because one clicked "agent" and the other clicked "complete."
Every vendor here has already priced for that. Microsoft splits licensed from unlicensed invocation and excludes computer-using agents. GitHub gives completions away and bills tokens. Google enforces a monthly credit on a weekly clock. OpenAI simply declines to publish. The models differ; the direction does not. The seat is the entry fee.
The organisations that get this right in 2027 will not be the ones that negotiated the best per-seat discount. They will be the ones who arrived at the renewal knowing their weekly actives, their token mix and the size of their index — three numbers, none of which is headcount.
Price the meter, not the seat. The seat is just the cover charge.
Continue Reading
- M&T Published 3 Copilot Counts. None Was Weekly Actives.
- Agentic AI Pricing: Don't Buy Consumption Without a Cap
- GitHub's Router Bills Every Leg. Find Your Break-Even.
- Glean vs Copilot vs Dust: Buy the Permission Model
- Google Bills Agent Memory Now. Revisions Count as Storage.
- One PR Billed 156M Tokens. Cap the Reads, Not the Rate.
- AI Observability Pricing: Same 10 GB, $49 or $930
- AI Gateway vs API Management: No Managed Gateway Caps Your Spend