Claude vs ChatGPT at work: if your data must be stored in Europe, only ChatGPT Enterprise documents that option; otherwise, ChatGPT is the better fit for business teams' everyday work, Claude for those who build and experiment, and measured usage, team by team, settles the rest. Of 151 companies that described their AI tools to Idun Group between 11 June and 28 October 2026, including 95 mid-sized and large companies, 55 have Claude in one form or another, 15 have ChatGPT and 9 have both: these are the tools they described to us, not market shares.
Claude and ChatGPT: which tools are we talking about at work?
Claude and ChatGPT each name two tools: a general-purpose work assistant, open to every business team, and a coding agent, Claude Code at Anthropic and Codex at OpenAI, built for technical profiles. Before comparing Claude and ChatGPT, you therefore have to say which family you are comparing: Claude's assistant with ChatGPT, or Claude Code with Codex.
| Work assistant | Coding agent | |
|---|---|---|
| At Anthropic | Claude (the chat and Claude Cowork, merged since September 2026) | Claude Code |
| At OpenAI | ChatGPT (merger with ChatGPT Work announced on 2 November 2026) | Codex |
| For whom | every business team | developers, data scientists, analysts, technical profiles |
| What it does | write, analyse documents, search, carry out a task with the company's tools | read and edit files on the computer, run commands, write code |
| Learning curve | low | significant |
Facts taken from Anthropic's help pages and the specialist press (checked on 17 November 2026).
The work assistant is the Claude or ChatGPT chat app, on the web, on desktop and on mobile. At Anthropic, the chat and Claude Cowork, the mode that works with your files, your emails and your internal tools, have been a single product since September 2026, in Team and Enterprise, wherever the administrator has turned it on. At OpenAI, the merger of the chat and ChatGPT Work, the mode that carries a task through from start to finish, was announced on 2 November 2026, according to TechCrunch: it is being tested, and Codex stays separate.
The coding agent acts on the computer: Claude Code and Codex read and edit files on the workstation, run commands and write code. They take time to learn, which the assistant does not, and they are used mostly by developers, data scientists, analysts and technically minded people.
The two families often live side by side in the same company: "The developers are on Claude Code. But even in marketing, for creative roles, we install Claude Desktop so they have access to the MCP connectors," a digital SME told us in July 2026. The line is blurring too: in 2026, at several of its clients, Idun Group saw technically minded business directors prefer the coding agent to the assistant. At one of them, the head of real estate, a former CIO, obtained an exception for Claude Code, which was reserved for the IT department, even though the group's tool for business teams was Dust. He uses it to analyse large volumes of data and build his reporting dashboards: with the coding agent, he does far more than with the assistant intended for him.
Claude Code or Codex: which one for technical teams?
For technical teams, Claude Code and Codex are almost level in 2026, according to Idun Group. Claude Code, Anthropic's coding agent, pulled ahead as soon as it came out, and many teams adopted it first; OpenAI's Codex has nearly caught up, and OpenAI is pushing it harder and harder. The choice remains open: "In the short term, the idea is to confirm a choice for the technical teams, between Claude and Codex," the head of data products at a mid-sized e-commerce company told us in October 2026.
Which is better, ChatGPT or Claude?
Claude has a slight edge over ChatGPT in 2026, too small to decide on its own. On LMArena's text leaderboard, where web users pick blind between the answers of two models, Anthropic's best model, Claude Opus 5.5, led OpenAI's best, GPT-5.6 Sol, by 22 points on 17 November 2026: its answer wins about 53% of head-to-head matchups. Both trail Google models, and the gap changes from one quarter to the next. Claude runs on Anthropic's Claude models, ChatGPT on OpenAI's GPT models: choosing one or the other also means choosing a family of models, but the gap between them does not settle the choice. For a company, the better of the two is the one its teams actually use, and that can be measured.
What decides before any measurement
Three requirements sometimes decide between Claude and ChatGPT on their own, based on the facts the vendors publish:
- Data stored in the European Union: of the four business plans, only ChatGPT Enterprise documents it as of 16 November 2026. Have OpenAI confirm the exact scope of its "EU" region before settling on it.
- Account management through SCIM, audit logs or a Compliance API: you need the Enterprise plan, from either vendor. SCIM automatically syncs accounts with the company directory.
- A budget you can forecast to the month: Claude Enterprise, and ChatGPT Enterprise under flexible pricing, bill by usage; the seat price is no longer enough to forecast the bill.
Why switch from ChatGPT to Claude?
From what Idun Group observes in 2026, companies that switch from ChatGPT to Claude do so first for their technical teams and for Claude Code, which many of them adopted early. The opposite move exists for business teams: "For everything that isn't technical, I advised them to move to ChatGPT rather than Claude: to me, it's cheaper and just as good," the head of AI at a digital SME told us in September 2026.
Claude or ChatGPT: which did the 151 companies Idun Group met choose?
Among the 151 companies that described their AI tools to Idun Group between June and October 2026, Claude is more common than ChatGPT, 55 against 15, but with few licences, and we often see it come in through the IT department and the technical teams; ChatGPT still has the image of a tool more accessible to business teams.
Source: exchanges between Idun Group, a French firm that audits AI usage in companies, and 151 company management teams (meetings, written exchanges, trade shows), from 11 June to 28 October 2026. "Claude" counts the assistant, Claude Code and Cowork; "ChatGPT" counts its plans, ChatGPT Team (the former name of ChatGPT Business) included, but not internal platforms built on OpenAI's models. The other 90 companies mostly have a Microsoft Copilot or an internal platform (53 of 90). These are the tools described, not market shares: our work on Claude may have brought more of its users to us.
Why is Claude seen as the developers' tool?
Claude is seen as the developers' tool because, at many of the companies Idun Group talks to, it came in through the IT department and the technical teams, while ChatGPT, opened to the general public in November 2022, is still seen as a tool for everyone. "I find Claude very developer-focused, and ChatGPT more made for the user, for ordinary mortals," a Claude administrator in the IT department of a social housing provider told us in September 2026.
Habits follow that image. In the browsing logs we analyse at our clients, ChatGPT is the AI site employees visit most outside the company's own tools, with personal accounts: this is shadow AI. That is no reason to choose ChatGPT, but it is worth knowing: employees equipped with Claude may keep opening ChatGPT on the side.
If Claude is already deployed in your company, our page on what Claude produces across your teams explains what its logs let you read, and what they leave out.
Why do companies move in small steps with Claude?
The companies Idun Group meets move in small steps with Claude because they fear the bill will soar with usage. We see them start with a few licences, often on the Standard seat of Claude Team and for the IT department: at the 13 companies that gave us their number of Claude licences in 2026, the median is 20, from 1 to 250. "On Claude, we are still at the case-by-case stage, with activations subject to budget," a programme manager at the French subsidiary of an international IT services company told us in September 2026.
Why does usage-based billing change the calculation?
Usage-based billing changes the calculation because the bill follows what the teams consume, no longer the number of seats: it is already the rule for Claude Enterprise at renewal, and an option in ChatGPT Enterprise. At a consulting and training firm with about forty Claude seats, Idun Group measured in 2026 that the consumption of the three heaviest users was worth 12 to 25 times their seat at the vendor's list price: a measure of consumption, not of value produced, detailed in our field report on Claude adoption.
Anthropic is pushing this shift, at least with some customers. In October 2026, a mid-sized industrial company with 322 Claude Team seats, across several workspaces, told Idun Group that Anthropic was offering it €1 million in free token credits to move to Claude Enterprise: a commercial offer made to this company, not a public price list. It had not said yes: in the long run, its IT manager reckoned, the vendor would come out ahead, not the company. "Future per-token billing may well call Claude into question," he added. The same month, the head of AI development tools at a large public group turned the question around: "When I make Claude Code available, people ask me what budget to plan for. I answer: what are your objectives?"
Claude Team and Enterprise, ChatGPT Business and Enterprise: what does each plan include?
In November 2026, Anthropic and OpenAI each sell two business plans: a plan for teams, Claude Team (2 to 150 people per workspace) or ChatGPT Business (2 to 200), and an Enterprise plan with more admin controls. All four include the vendor's coding agent. They differ mainly in what is paid for by usage and in what the administrator can see.
| Claude Team | Claude Enterprise | ChatGPT Business | ChatGPT Enterprise | |
|---|---|---|---|---|
| For whom | 2 to 150 people | large companies, 20 seats minimum | 2 to 200 people | companies operating at scale, on quote |
| The seat | Standard, or Premium with five times more usage; limits set member by member | access only, with no usage included | Standard, or Premium with more usage and no five-hour limit | seats with limits, or flexible pricing |
| Beyond the seat | prepaid usage credits, under spending caps | all usage at API rates, under caps per organisation and per person | credits to buy | credits or tokens, with no fixed limit under flexible pricing |
| Single sign-on and accounts | SSO; no SCIM | SSO and SCIM | SAML SSO; no SCIM | SAML SSO and SCIM |
| Connectors | Google Drive, Gmail, Google Calendar, GitHub, Microsoft 365, Slack | the same | Google Workspace, Slack, GitHub, Microsoft 365, among others | the same, plus a connector registry |
| Coding agent | Claude Code included | Claude Code included, its usage paid separately | Codex included | Codex included |
| What the administrator sees | usage analytics; no audit logs or Compliance API | usage analytics, audit logs, Compliance API, Analytics API | analytics dashboard, basic user analytics; no Compliance API | user analytics, Compliance API |
Facts taken from Anthropic's and OpenAI's plan and help pages (checked on 16 November 2026).
Seat prices are on the plan pages of Claude and ChatGPT, read on 16 November 2026. We do not copy them here: they change, and the budget depends mostly on what the seat includes and what is paid beyond it.
In Claude Enterprise, every request is paid at API rates, meaning per token consumed, under the caps the administrator sets; contracts signed with the older seats, Chat or Chat + Claude Code, move to this model at renewal. ChatGPT Enterprise leaves a choice: seats whose limits are, for most Codex and ChatGPT Work features, those of the Plus plan, or flexible pricing with no fixed limit. From December 2026, ChatGPT and Codex will share a single usage limit in ChatGPT Business. Claude Team, for its part, sets limits member by member: anyone who reaches theirs can only continue with prepaid credits, even if other seats in the team sit idle.
Once the seats are handed out, there is still ChatGPT Enterprise adoption beyond the early enthusiasts.
What happens to your data in each plan?
None of the four Claude and ChatGPT business plans is used by default to train the vendor's models. The differences lie in retention and storage location: ChatGPT Enterprise offers data residency in ten regions, the European Union among them, while Anthropic states that data from its commercial plans is stored in the United States.
| Claude Team and Enterprise | ChatGPT Business and Enterprise | |
|---|---|---|
| Model training | No, by default. Exception: feedback the user sends, kept for up to five years; workspace owners can turn off this sending | No, by default, in both plans |
| Retention | Conversations stay in the product; a deleted conversation leaves Anthropic's servers within 30 days. Custom retention in Enterprise only, 30 days minimum | Custom retention policies in Enterprise |
| Location | Processing possible in the United States, Europe, Asia and Australia; storage in the United States. Usage-billed Enterprise: US-only inference option | Enterprise: data residency in ten regions (United States, European Union, United Kingdom, Japan, Canada, South Korea, Singapore, India, Australia, United Arab Emirates). Business: no data residency |
| Certifications | SOC 2 Type I and Type II, ISO 27001:2022, ISO/IEC 42001:2023, for commercial plans | SOC 2 Type 2 for Business and Enterprise; ISO 27001, 27017, 27018 and 27701 for Enterprise |
Facts taken from Anthropic's privacy pages and OpenAI's plan pages (checked on 16 November 2026).
Claude models are also available in European regions of Amazon Bedrock, including Frankfurt and Paris, according to AWS documentation (checked on 17 November 2026). That is a different product, API access to the models for building your own tools, without the Claude app or its consoles. Two points remain to settle with the data protection officer: personal subscriptions, Claude Pro or Max, ChatGPT Plus or Pro, paid through expense claims, are not covered by this table; and both Claude Enterprise and ChatGPT Enterprise let the company manage its own encryption key.
Can you connect ChatGPT to Claude, or move from one to the other?
Not directly: in November 2026, Claude and ChatGPT have no native integration with each other. To use them together, you need a platform open to both vendors' models, such as Dust, a French platform for agents shared across the company. The open formats they share, on the other hand, let you move from one to the other without getting locked in. MCP (Model Context Protocol) connectors, which link the assistant to your applications and databases, and skills, reusable instructions and procedures that an agent loads when the task calls for them, are managed by the Agentic AI Foundation (AAIF), a foundation under the Linux Foundation created on 9 December 2025 by Anthropic, OpenAI and Block. Anthropic handed it MCP, which it had launched in November 2024, then the Agent Skills format on 22 October 2026; more than 70 platforms, ChatGPT among them, had already adopted Agent Skills, according to Anthropic.
Claude, ChatGPT and their coding agents will keep overtaking and copying each other, and other players will arrive. So build your connectors in MCP and your skills in the Agent Skills format, keep them in a repository you own, and document them: they can then follow the company to another tool. A plugin is only portable through the MCP connectors and skills it contains. At Atos, one team goes further: "We put all our skills in OpenCode, so we can switch if needed," a technical director at the group told us in October 2026. OpenCode is an open-source coding agent, under the MIT licence, that works with models from more than 200 providers, including models installed locally (checked on 17 November 2026).
Is there a sovereign alternative to Claude Code and Codex?
A sovereign alternative partly exists, but in 2026, at the companies Idun Group meets, none has yet taken the place of Claude Code or Codex. The French option is Mistral Vibe CLI, Mistral AI's command-line coding agent, released under the Apache 2.0 licence on 9 December 2025 with the Devstral 2 model. OpenCode becomes a sovereign option when you run it with a model hosted in Europe or installed locally: open source does not mean sovereign. Idun Group's advice: accept dependence on an American vendor for now, while keeping connectors and skills portable, so you can move to a sovereign tool the day one does the same work.
Can you tell, after 90 days, which one is actually used?
After 90 days, the Claude and ChatGPT consoles show who uses each tool, how much, and increasingly for which kinds of tasks. To know whether that work served the team's goal, you have to set these figures against its results and against interviews, because no console compares the two tools or sees personal subscriptions. OpenAI says as much in its own documentation: a message count does not tell you how much work the team finished. We use 90 days, the depth of Claude's spend export.
What the Claude console shows
- Usage analytics in Claude Team and Enterprise give weekly active members, assigned seats, a breakdown by group and by member, the connectors used, spend by model and an estimate of time saved, calculated by Anthropic.
- The spend export gives one row per person and per model, over 90 days at most; in Enterprise contracts still billed by the seat, only the spend beyond the seat.
- For Claude Code, a dashboard gives lines of code accepted, the acceptance rate, daily active users and sessions; when linked to GitHub, the pull requests that contain code written with Claude Code.
- In Enterprise only: an Analytics API, audit logs, the Compliance API and, since 28 October 2026, "smart reports". To produce them, Claude reads a sample of a team's conversations, over 28 days at most, and draws out the workstreams, deliverables, spend and blockers; these reports are not meant, Anthropic says, for evaluating individuals. Built-in surveys also ask members what Claude has brought to their work.
What the ChatGPT console shows
- Business and Enterprise have an analytics dashboard and basic user analytics.
- The Insights view sorts ChatGPT Work and Codex usage into categories and tasks, with the messages, credits and active users for each, sometimes on a sample.
- An Analytics API gives daily usage of ChatGPT, Codex and ChatGPT Work.
- In Enterprise only, the Compliance API gives access to conversations and agent activity. OpenAI states that it is not a productivity dashboard (checked on 16 November 2026).
Should you choose, or keep both?
Keeping both Claude and ChatGPT makes sense when each serves different teams or different tasks, for example one vendor's work assistant for business teams and the other's coding agent for the IT department; if the same people have two seats for the same tasks, one of them is too many. Nine of the 151 companies Idun Group met have both, and some IT departments are trying to consolidate: "We chose Claude. Other houses in the group went for ChatGPT instead. The IT department's idea is to standardise and move to a single model," a technical employee outside the IT department at a wine and champagne house told us in October 2026. The decision is made tool by tool, on the cost per active user, not on the seat price. The same reasoning applies between Copilot and ChatGPT.
How do you calculate the cost per active user?
A tool's cost per active user is its total cost over the period, divided by the number of people who actually use it. Put your own figures in.
- The total cost: the seats, at your contract price, plus everything paid on top: prepaid usage credits for Claude Team, usage at API rates for Claude Enterprise, credits bought for ChatGPT Business and Enterprise. Add the personal subscriptions reimbursed for the same tool (Claude Pro or Max, ChatGPT Plus or Pro).
- Active users: the unique people active over the period according to the console, not the seats assigned, with each console's definition noted alongside. Count personal subscriptions once their activity is confirmed, without double-counting anyone who also has a seat.
The calculation is: (seats + usage billed on top + personal subscriptions) ÷ active users. An idle seat raises this cost without the bill moving.
Extend, keep or cut
You decide whether to extend, keep or cut on this cost, read alongside what the tool returns: a tool that is expensive per active user can carry a task that pays off, and a cheap one can duplicate another. Idun Group's method to measure what each assistant returns details this calculation from actual usage.
- Extend the tool that costs little per active user and whose teams describe a gain on their main tasks.
- Reassign or cut seats without regular use over two quarters, that is, two consecutive 90-day readings, and the duplicate tool when the same teams do the same tasks with both.
- Train before cutting: a little-used seat sometimes comes down to features never tried. According to OpenAI's report "The State of Enterprise AI" (December 2025), 19% of ChatGPT Enterprise users active in the month had never used data analysis.
The checklist to pass on before renewal
Gather five measures, over the same 90-day period if possible, before you choose, renew or cut. No single console gives them all: note what is missing rather than estimate it. The data or AI manager can then pass them on to management as they are.
- Seats paid and seats active, per tool. Where to read it: Claude's usage analytics, ChatGPT's analytics dashboard.
- The teams that use both tools, and what each does with one and with the other. Where to read it: the group breakdown of both consoles, cross-checked.
- The three dominant use cases per team, taken from the consoles' categories and confirmed by the team. Where to read it: Claude Enterprise's smart reports, ChatGPT's Insights view.
- The cost per active user of each tool, calculated as above. Where to read it: Claude's spend export, the credits billed by OpenAI, expense claims.
- What the interviews say, set against the measurement, each finding marked "measured" or "reported". Where to read it: in the interviews; Claude Enterprise's surveys give a first, reported part.
Gathering them means reading two consoles that do not talk to each other and asking the teams: your team can do it with this checklist. It is also the work of Idun Group, a French firm that audits AI usage in companies: its audit reads actual Claude and ChatGPT usage, team by team from the exports your consoles already produce, and looks at usage, not individuals.




