BusinessCost Playbook9 min readPublished

Copilot vs ChatGPT vs Claude at Work: Microsoft’s Cost Claim

Microsoft’s $30, $81 and $205 figures describe the same modelled month of Office work. The seat prices, included usage and models chosen for that work differ substantially.

DA
Digital Applied
PublishedNovember 19, 2026
Read time9 min

Microsoft puts the monthly cost of an everyday AI user at $30 with Copilot, $81 with ChatGPT Enterprise and $205 with Claude Enterprise. Before taking those numbers to finance, ask what happens when an employee opens a spreadsheet. In Microsoft’s comparison, that work was included in the Copilot seat. For the other products, the cost model added consumption to a seat fee. That distinction explains why the $30 can hold steady while the other two totals rise.

Microsoft commissioned and paid for the research, conducted the tests and calculated costs using October 2026 pricing. GigaOm helped design the research, validated the method, results and cost analysis, and co-authored the report, Demystifying the TCO of Enterprise AI. Microsoft published its account on November 16. The research gives buyers an unusually specific workload to examine, provided they keep the model choices and pricing assumptions with the headline. Microsoft’s study summary and link to the report

Key takeaways
  1. 01
    Each total has a different price basis.Figure 1 uses a $30 list-price Copilot seat and a $20 list-price Claude Enterprise access seat. The $20 ChatGPT Enterprise seat is an assumption, as is the use of $0.04 credits for that comparison.
  2. 02
    The everyday result is about included work.The 100 monthly chat and Office tasks add no modelled usage charge to Copilot. At 140 monthly tasks, including 20 agentic workflows, Copilot rises to $96.
  3. 03
    The model selection was asymmetric.Microsoft used Auto in Copilot, GPT-6 Sol in ChatGPT and Opus 5.5 in Claude. The tests supplied comparable work, but each product did not use the same routing strategy.
  4. 04
    Price the product your staff would buy.Business or Team seat prices, an enterprise access fee and credits for extra usage describe different purchases. A public small-team price cannot establish your enterprise monthly total.

01 — The figuresWhat does the monthly total buy?

The everyday profile completed 25 tasks a week for four weeks. The advanced profile completed 35 a week, five of them agentic workflows. The table shows the modelled cost per person for each four-week month. Its middle column is the study’s seat basis. Subtracting that from each profile’s total makes the implied usage component visible. Figures 1 and 2 of the report give the seat basis and the rounded totals.

October 2026 model, USD per user per month. Parentheses show usage above the study’s seat fee, calculated from its published totals.
ProductSeat in the studyEveryday · 100 tasksAdvanced · 140 tasks
Microsoft 365 Copilot$30 list$30 ($0 usage)$96 ($66 usage)
ChatGPT Enterprise$20 assumed$81 ($61 usage)$149 ($129 usage)
Claude Enterprise$20 list$205 ($185 usage)$299 ($279 usage)

For Copilot, the $66 difference comes from modelled Cowork usage, because the ordinary chat and Office work in both profiles is included. For the other two products, subtracting the everyday total from the advanced total would give a misleading agent price: the mix of everyday tasks changes too. The advanced user does fewer plain chats and more work inside Word, Excel and PowerPoint. The source gives a combined bill for that different mix.

If a buyer already pays for the qualifying Microsoft 365 base licence, $30 can be a meaningful incremental seat cost. Someone comparing complete software estates needs to price the base separately. The study’s totals also omit the labour of administering or reviewing these products. A decision that depends on permission management, deployment time or how long a human spends fixing an answer needs evidence beyond the monthly table.

02 — Figure oneThree meanings of a seat

For Microsoft, Figure 1 uses the $30 Copilot list price and $0.01 per Copilot Credit for Cowork. Everyday chat and work in the three Office applications were included. For Claude Enterprise, it uses a $20 access fee explicitly described as list price, plus recorded usage converted to list-price equivalents. For ChatGPT Enterprise it assumes a $20 monthly seat, then prices recorded credits at $0.04 each. The report says the ChatGPT simplification was informed by its test tenant.

The difference between the two $20 figures matters. Anthropic publishes a price for an Enterprise access seat and separately meters consumption. OpenAI publishes custom Enterprise pricing. Reading the study as though both had sold Microsoft the same public $20 offer would give a procurement team a false starting point. Nor can you transfer the price of a credit between the vendors: the unit buys different things and the services count different events.

A finance lead can use the assumptions to locate the point of disagreement. The ChatGPT seat starts $10 below Copilot in the study. At the assumed four cents per credit, an additional 250 credits would use up that gap if each credit incurred a marginal charge and Copilot stayed at its included seat price. A pool of credits already purchased under a contract may make the marginal charge different. Work out what is included, consumed and invoiced before extrapolating the $81.

03 — Work sampledWhich 37 prompts drove the result?

Microsoft used 12 distinct prompts in chat, seven in Word, seven in Excel, six in PowerPoint and five in Cowork. The prompts were repeated across profiles and fresh sessions. In the everyday week the model counts 12 chat tasks, four Word, five Excel and four PowerPoint. The advanced week counts eight chat, eight Word, eight Excel, six PowerPoint and five agentic workflows. These are two designed patterns of work. The report does not establish that an average employee at your organisation performs either one.

Each prompt ran up to ten times per product against the same fictional company setup and data. Human evaluators used prewritten pass criteria, including selecting the correct sources, preserving facts and completing the requested deliverable. A run had to meet every criterion to pass. The authors describe quality across the selected workload as broadly comparable. That permits a cost comparison for that workload, while leaving substantial room for individual tasks to favour one product.

Consider an analyst who spends most of the week in an existing workbook and a coordinator who mostly searches meeting notes. A combined score can hide a serious problem in a spreadsheet formula or a citation. Ask their managers for one recent accepted deliverable from each group. The reviewer should be able to say exactly what would cause it to be rejected. If a workbook’s totals must recalculate when a source value changes, a screenshot with the right number is an inadequate success criterion.

04 — RoutingAuto had an advantage the others did not exercise

For chat and Office work, Microsoft ran Copilot in Auto. Figure 1 specifies Auto at medium effort for Copilot Cowork. Claude Enterprise ran Opus 5.5, using medium effort for chat and Cowork and extended thinking in the Office add-ins. ChatGPT Enterprise used GPT-6 Sol at medium for its chat, Office add-ins and ChatGPT Work. This was October’s model configuration. GPT-6.2 Sol, released November 17, was outside the comparison.

Auto can route a quick rewrite differently from an analysis that needs more reasoning. Running the same premium Claude model for both jobs charges every request at that model’s usage rates. For the agentic workflow cost estimates, Microsoft read session cost inside Copilot Cowork. It gathered token information for Claude Cowork and ChatGPT Work via Claude Code and Codex respectively, then converted usage using public rates. Those methods can inform a budget. They are different from comparing three settled monthly invoices.

An organisation that expects staff to accept the defaults may care about precisely that advantage. Another may configure one model for cheap text tasks and reserve a more expensive one for the few cases that fail review. Try both policies on the same documents. Compare the cost of a result that passes, including failed attempts and reviewer time. If a cheaper route misses a required source or breaks a formula, its low token price does not settle the choice.

05 — Buying todayWhat the vendors sell in November

As of November 19, Microsoft lists its US enterprise Copilot add-on at $30 per user per month paid yearly, with a qualifying Microsoft 365 licence required. Its partner guidance announces new offers from December 1, at the same price, that can be purchased without that base-subscription purchase prerequisite. This changes the order a buyer can place. It does not make the October benchmark a comparison of complete office suites. Microsoft enterprise pricing · Microsoft’s December SKU change

Microsoft’s small-business offer is a separate purchase. For eligible existing Microsoft 365 Business customers, the US Copilot Business add-on has a regular $21 per user monthly price on annual commitment. A first-year promotion lowers it to $18, offered through December 31, 2026. A bundled business plan with an underlying Office subscription has a different price again. Applying that promotional $18 to the enterprise study would mix the customer segment, the product and the price period. Copilot Business offer and terms

OpenAI lists its US ChatGPT Business Standard seat at $20 monthly with annual billing or $25 with monthly billing. Its Premium seat is $100 or $125 on the same bases. Business and Enterprise customers can buy credits for more access. Enterprise pricing is custom, so the Business Standard sticker price cannot confirm the study’s ChatGPT Enterprise assumption. OpenAI business pricing

Anthropic lists the same four US seat prices for Claude Team Standard and Premium. On Claude Enterprise it posts $20 per seat per month, billed annually with a 20-seat minimum, plus usage at API rates. That structure explains how a $20 Claude access fee could yield a $205 modelled month. A Team buyer receives plan usage with a seat and should compare the applicable inclusion and additional usage terms before substituting the Team price into a spreadsheet. Anthropic’s team and enterprise pricing

06 — November releaseThe Office apps are now inside Copilot

On November 17, Microsoft made Word, Excel and PowerPoint generally available inside Copilot. That can remove a handoff: an employee who starts by discussing a report can continue into the document, workbook or deck in the same experience. Microsoft says these everyday Office tasks are included in its licensed Copilot experience. This release came after the October test. It changes what a buyer can try now, without changing what Microsoft measured then. Microsoft’s Ignite release record

Autopilot is a different cost question. Microsoft is rolling this persistent agent out through its Frontier programme, and its licensing guide gives preview consumption rules. The US pay-as-you-go rate is $0.01 per Copilot Credit. The guide lists 100 credits per agent per month for the persistent account and resources, plus usage as the agent works. At that unit rate the recurring component is $1 before activity. The guide describes Frontier preview rules and planning estimates. Final generally available Autopilot pricing remains open. Microsoft’s Copilot Credits licensing guide

Microsoft also put shared Copilot Workspaces into Frontier and introduced Agent 365 Business Analytics there to connect spend with work performed. Neither is a measured benefit in the October cost totals. For a pilot, pair whatever usage information you receive with a review of the actual output. Opening a spreadsheet successfully or logging agent activity costs money even when the reviewer sends the file back. Microsoft’s November 17 Copilot announcement

07 — ProcurementGive each product the same job

Begin with an explicit decision. Are you adding an assistant to an estate you already own, or buying the core office tools as well? What share of employees will receive seats but use the assistant rarely? Give that group its own row. Microsoft’s everyday user carries out 100 tasks a month, and a buyer cannot assume that all allocated seats reach that level.

For scale, suppose 80 users matched the study’s everyday profile and 20 matched the advanced profile exactly. Multiplying the published totals yields $4,320 a month for Copilot, $9,460 for ChatGPT and $22,380 for Claude. This calculation inherits every October price, routing choice and task assumption. Its purpose is to give finance an editable baseline. Change the active headcount, included credit pool, selected Claude model or paid Copilot workflows and recalculate.

Ask for the prompt list through your Microsoft account team if you want to reproduce the published workload. To judge your own teams, supply the same representative files and access to all three products, record the model and effort, then grade the deliverables before inspecting the usage bills. The Copilot Cowork guide covers the agentic work, and our AI spend controls comparison covers how budgets are enforced. A brief for an AI adoption pilot can specify both. The earlier Autopilot overview provides context for the newly metered work.

Existing staff work mostly in Word, Excel and PowerPoint
Measure included Office use and the volume of separately metered agent work.
Price the add-on
Your AI contract pools usage across users
Check seat fees, credits already committed, overage rates and how many people actually use them.
Price the pool
A mistake is costly to repair
Apply a reviewer’s checklist to repeated runs and include the cost of corrections.
Price the accepted result
Decision

Put the task mix beside the price

Microsoft’s research makes a credible case for included Office work in an organisation already using Microsoft 365. It also identifies exactly what can change the answer: the seats purchased, models selected, tasks completed and credits already covered. Ask each vendor to price that mix and have the same reviewers decide which outputs they would accept.

Digital Applied

Evaluate an assistant on an actual Office workflow

Bring a workbook, a document and a repeatable agent job to compare quality and cost under the terms available to your team.

Seat assumptionsUsage chargesReview criteria
Evaluation

For a procurement trial

  • →Realistic weekly tasks
  • →Current entitlements
  • →Acceptance checklist
Answers

Frequently asked questions

Microsoft commissioned, funded and conducted the testing and cost analysis. GigaOm helped design the research, validated the methodology, results and costs, and co-authored the report, Demystifying the TCO of Enterprise AI.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source