Claude vs OpenAI: plans and platforms

Two vendors, four products

The single most useful thing to hold in your head is a two-by-two: each vendor sells an app product bought per seat, and a developer product billed per token. Four products, two shapes. Almost every procurement question resolves once you know which of the four squares you are standing in. The two developer squares each have their own explainer - What is the Claude Platform? and What is the OpenAI Platform? - and each vendor's seat-versus-token split has its own article; this one is about how the four squares line up.

Read the columns and the vendors look nearly identical. The interesting material is in the rows: how each one subdivides an organization, who can do what inside it, and - the real differentiator - how much traffic is allowed between the app side and the dev side.

The translation table

Teams that use both vendors lose more time to vocabulary than to technology. These are the pairs worth memorising.

The third row is the one that bites. "Workspace" means the team account at OpenAI and a project-like subdivision at Anthropic. Two people using the word in the same meeting will be describing completely different objects - one a company-wide container for seats, the other one of a dozen buckets holding API keys.

Account structure

Both developer platforms solve the same problem - keep staging from spending production's budget - and they solve it the same way, with a named subdivision inside the organization that owns keys, members, and limits.

Practically: if you already run a tidy OpenAI account with one project per environment, the Claude Console asks nothing new of you - create one workspace per environment and carry on. The one habit to add is that Anthropic will not let you re-home a key, so name and scope workspaces before you start minting credentials.

Walls and doors

Both vendors put the subscription and the API in separate accounts with separate bills. So there is exactly one question that separates them: can money you already pay for seats ever pay for API work? At OpenAI the answer is no, in every case. At Anthropic the answer is yes, in four specific cases that Anthropic built on purpose.

That is the whole section. The picture below is the same idea: two chambers on each side - seats on top, tokens underneath - and the only difference is whether anything can flow between them.

Neither design is better in the abstract, and the trade is easy to state. OpenAI's walls give finance a clean split - a fixed headcount cost and a separate variable cost, with no way for one to leak into the other. Anthropic's doors keep a developer working when a limit hits, at the price of an invoice that mixes both and that somebody has to be able to explain.

The mistake to avoid is assuming the vendor you know behaves like the one you don't. If you came from OpenAI, note that at Anthropic a plan limit is not a spending ceiling. If you came from Anthropic, note that at OpenAI there is no overflow route at all - the limit is the limit.

Admin and governance

On the app side the two are close: both offer a team tier with central seat management and business data terms, and an enterprise tier above it that exists for SSO, provisioning, audit logs, and compliance rather than for better model access.

On the developer side, Anthropic's role model is the more granular of the two. The Console distinguishes members who may only use the Workbench, members who may also use Claude Code, developers who manage keys, billing roles, and admins who manage users - and the elevated tiers reach into every workspace whether you add them or not. That granularity is useful for separating a data scientist experimenting in the Workbench from an engineer holding production credentials.

Both vendors now expose administration through an API, which is what makes onboarding and offboarding automatable rather than a checklist. If you are standing up either platform for more than a handful of people, wire that in early - the alternative is discovering six months later that a departed contractor's key is still live.

One rule applies identically to both: keys belong to programs, not to people. Scope them to the narrowest subdivision that works, and never let one reach a browser. That, and the spend limits, are the two controls that actually prevent incidents.

The coding tools

Both vendors ship an agentic coding tool, and this is where the wall-versus-door distinction stops being an accounting curiosity and starts affecting your engineers' day.

Anthropic includes Claude Code on paid Claude seats, sharing the plan's usage limits - so a developer with a seat can work in the terminal without anyone provisioning an API key, and can optionally spill over onto API-rate credits when the plan runs dry. OpenAI's Codex is available through the ChatGPT plans that include it, with API-based routes available separately; the two funding sources stay distinct.

The operational question is the same on both sides and it is not "which tool is better": it is which budget a developer's coding usage lands on, and what happens when they hit the ceiling. On the Anthropic side there is a documented path onward, which is convenient and needs a decision made in advance. On the OpenAI side the ceiling is the ceiling until someone provisions API access, which is inconvenient and needs no decision at all.

For a like-for-like look at the tools themselves rather than their billing, see Copilot vs Claude Code.

Choosing, and running both

Most engineering organizations end up with accounts at both vendors, so the useful question is rarely "which one" - it is how to keep four accounts legible.

• Name an owner per account, not per vendor. Four products, four owners: two app-side subscriptions belong to whoever owns headcount, two dev-side organizations belong to engineering.
• Mirror your structure across both platforms. One project per environment at OpenAI, one workspace per environment at Anthropic, identical names. Usage reports then answer the same question on either side.
• Set spend limits before the first production call, on both. They are the only control that works while everyone is asleep.
• Decide the Claude Code funding question explicitly - plan limits or Console credits - and write it down. It is the one question OpenAI's design answers for you and Anthropic's does not.
• Expect four line items in finance, two fixed and two variable. Approving them from one pooled AI budget is how variance gets invisible.

And the framing that survives every pricing change either vendor makes: a seat is a person, a key is a program. Every question about which product you need, which bill something lands on, and who should own it comes back to that one distinction - and it is the same distinction on both sides of the aisle.