Claude Enterprise agreement clauses that decide your cost. Renewal, overage, models and data.

A Claude Enterprise agreement is a $20 seat plus usage billed at API rates. What you pay over the term depends on four clauses: the renewal cap, the overage rate, model substitution and data use.

On most Claude Enterprise order forms the seat fee is the smaller of the two charges. The larger sums sit in usage and in the terms that apply at renewal, and a handful of clauses decide both. This page explains each one and the wording to ask for.

This guide is for procurement, legal and AI platform leaders papering a Claude deal. Read it alongside our Claude enterprise licensing guide, which covers the commercial model, and this separate guide to Anthropic Claude negotiations. Our GenAI Practice negotiates the commercial and the legal terms together, because they interact.

How is a Claude Enterprise agreement priced in 2026?

A Claude Enterprise agreement has two charges. The seat fee is $20 per user per month, billed annually, and it includes no usage at all. Every chat, file analysis, Cowork task and Claude Code session is billed on top at Anthropic's standard API rates, pooled across the whole organization.

• Seat minimums. Anthropic's help center lists 20 seats for self serve Enterprise and 50 seats for sales assisted contracts.
• Billing. Sales assisted customers are invoiced monthly in arrears for actual consumption. Self serve customers buy credits upfront, and usage stops when the balance runs out.
• Legacy seats. Organizations still on the older Standard and Premium seats, or on chat and Claude Code seat bundles, move to the single usage based seat at renewal.
• Spend limits. Admins can cap spend for the organization and for individual users. A cap halts consumption when it is reached, but consumption up to that point is still billed at the contract rate.

A seat fee is easy to forecast and approve. The usage commitment behind it, and the rate charged for usage above that commitment, is where most of the annual spend sits. Both are set by the clauses in the table below.

What have we seen in recent Claude and GenAI enterprise contracts?

We reviewed roughly 15 to 25 enterprise GenAI contracts between 2024 and 2025, Anthropic Claude deals among them. The headline price was rarely what cost the buyer money. The clause structure was, and three patterns came up again and again.

• No renewal cap. Second term pricing reset 20 to 40 percent higher than the first term.
• Uncapped overage. Usage above the commitment was billed at 1.5 to 2 times the committed rate.
• Pricing tied to one model. Deals written around a single model name had to be renegotiated mid term when that model was deprecated.

The sections below take the clauses in the order they tend to cost money, starting with renewal.

How should the renewal cap and price hold read in a Claude contract?

The renewal clause should put a percentage ceiling on any increase and state how long current rates hold. Most of the cost of a Claude deal lives in the second term, and a first term discount means little if the renewal resets to list.

Anthropic publishes its plans and API rates on the Anthropic pricing page, and it has changed per token prices as new model generations launched. Our record of Anthropic pricing and model changes tracks those shifts. The enterprise contract is the only place where a price hold and a cap become binding, so state both there.

Why cap the uplift to a percentage?

A percentage cap gives finance a number to forecast. Renewal tied to undefined list pricing hands the vendor an open variable that can reset your budget every term.

Write the cap so it covers the seat fee, the usage rates and your discount off API list, since a cap on one line invites an increase on another. Our uplift cap redline has sample language.

How long should the price hold run?

• Single year hold. Simple to agree, but it exposes you to an annual reset.
• Multi year hold. Locks the rate, which is worth most when usage is scaling fast.
• The trade. A longer hold is worth a tighter commitment only if the cap is real and applies to every line on the order form.
• Resizing. On a multi year hold, add an annual right to resize the usage commitment, so the rate protection does not lock you into a volume set before you had real data.

Our note on price hold clause language shows how to word the rate schedule so it survives a change in Anthropic's public list.

What do the consumption and overage clauses control?

They control what you pay for usage above your committed volume. Claude Enterprise deals usually pair a committed annual spend with usage billed at API rates, and the negotiated discount sits on the committed portion. The overage clause decides whether that discount survives when adoption grows faster than planned.

How do you avoid uncapped overage?

Set the overage rate in the contract, at or close to the committed rate. Add a notice threshold, for example at 80 percent of the annual commitment, so a spike is flagged before it bills. Then add the right to true up the excess into the committed tier rather than paying spot rates on it.

Admin spend limits are an operational control. Anthropic notifies users at 75 and 95 percent of a limit, and at the limit they are blocked until an admin raises it, a cost of its own if Claude Code or Cowork sits inside a delivery workflow. We cover the mechanics in how to cap an AI consumption overage.

How do you size the committed volume?

Size the commitment to observed usage from a pilot, not to a projection. A commitment built on ambition funds volume you may never reach. Use the admin console spend report, which exports per user and per model token use as CSV, or the Analytics API, over at least three months of steady use.

Sign before usage settles and you commit against the vendor's forecast. Our guide to pilot timing covers when to open the commercial discussion.

What does the overage clause cost in a worked example?

Say you run 1,000 Claude Enterprise users. The seat fee is 1,000 × $20 × 12, or $240,000 a year. You commit $500,000 of annual usage at a negotiated rate, and actual usage comes in 30 percent above plan: $150,000 of extra consumption valued at the committed rate.

The overage clause alone changes year one by up to $150,000. Renewal compounds it. If the second term is priced off the $650,000 usage run rate with no cap and resets 20 to 40 percent higher, usage costs $780,000 to $910,000 a year. A 5 percent cap would hold it at $682,500.

How should model substitution terms be written?

Tie the commercial rate to a capability tier rather than to a named model, and state what happens when that model is retired. Anthropic retires models on a published schedule, and a contract priced against one model name can reopen the deal mid term.

Anthropic's model documentation lists the current lineup and its lifecycle stages: active, legacy, deprecated and retired. The deprecation policy promises at least 60 days' notice before a publicly released model is retired. On April 14, 2026, for example, Anthropic announced that Claude Sonnet 4 and Claude Opus 4 would be retired on June 15, 2026.

How do you handle model versioning?

• Define tiers. Name the capability tiers you buy, such as the Opus, Sonnet and Haiku families, and price each tier. List the current model only as an example of the tier.
• Successor pricing. The successor in a tier costs no more per token than the model it replaces, or your discount carries over as a percentage off whatever list then applies.
• Migration window. For models in production, ask for more notice than the policy minimum, matched to your own release and testing cycle.
• No repricing on retirement. State that a retirement does not reopen the commitment or the discount.
• Reallocation. Let committed spend move between tiers, so a shift from Opus to Sonnet does not strand budget. Our note on swap and reallocation rights has the wording.

Why we advise against locking a model version into the contract

The usual advice is to lock the model version in the contract so pricing and behavior stay stable. We disagree, because the lock does not outlast the vendor's lifecycle. A retired model stops answering requests whatever the contract says.

In the deals we reviewed that were priced against one model name, retirement forced a renegotiation mid term. Put the tier and the successor price in the contract, and keep version pinning in your own application testing.

What data protections belong in a Claude Enterprise contract?

Write three protections into the contract: no training on your prompts and outputs, a defined retention window, and deletion on termination. Anthropic's public position covers the first already, but a policy page can be revised during your term and a signed contract cannot.

Anthropic's pricing page says Team and Enterprise plans include no model training on customer content by default, and "by default" is the phrase to close. Its privacy center adds that, where your contract permits it, Anthropic may anonymize your data for research or statistical purposes and keep it longer. Decide whether you allow that.

What do Anthropic's default retention rules leave open?

• Chats kept indefinitely. Enterprise chats and projects are kept until deleted, unless an owner sets a custom retention period. The shortest custom period is 30 days.
• Deletion lag. Deleted conversations leave chat history at once and leave backend storage within 30 days.
• Flagged content. If a chat is flagged under the usage policy, inputs and outputs can be kept for up to 2 years and safety classification scores for up to 7 years.
• Feedback. When a user submits a thumbs up, thumbs down or bug report, the associated data is kept for 5 years.
• Zero data retention. Anthropic offers zero data retention on the API by agreement, but its documentation excludes the Team and Enterprise product interfaces. The one exception is Claude Code used through Enterprise with zero data retention enabled.
• Feature gaps. Custom retention does not cover every surface. Features built on Claude Code on the web are among Anthropic's listed exclusions.

What should the data clause say?

Define prompts, outputs, uploaded files and connector data as your content, and state that none of it trains any model. Exclude or narrow the anonymized research use. Set the retention period in the order form as well as the admin console, and agree a rule on user feedback.

Require deletion within a fixed number of days after termination, with written confirmation. List the subprocessors and regions you accept. Anthropic offers US only inference, priced at 1.1 times standard input and output token rates, so if residency matters, contract for both. Our guide to AI data governance clauses covers IP and privacy wording in more detail.

What will Anthropic's account team say, and how should you answer?

Expect the same few positions in most Claude negotiations. Each one has a precise answer that keeps the clause open.

• "Renewal follows our list pricing at the time." Ask for a named percentage cap covering seats, usage rates and discount level. If list is not expected to move much, the cap costs Anthropic little to give.
• "Spend limits in the console give you cost control." A limit decides when usage stops. The rate for usage above the commitment still needs to be in the order form, along with the true up right.
• "We never train on commercial data. It is in our terms." Then writing it into the order form, with the anonymized research use removed, should be an easy yes.
• "You will want our newest models, and they are priced competitively." Agree, and write that into the contract as successor pricing within each tier.
• "A bigger commitment earns a bigger discount." Ask for the discount at two or three commitment sizes and pick the one your pilot data supports. The true up right covers growth.

How do you keep your options open for the next Claude renewal?

Your position at renewal depends on how easily you could move. A buyer that can show a priced, credible alternative gets a better renewal offer than one who cannot.

• Keep assets portable. Store prompts, system instructions and evaluation sets in your own repository, outside the vendor's projects.
• Document an exit path. Know what switching would take: which integrations, which Claude Code workflows, and how many weeks of testing.
• Benchmark before signing. Price at least one credible alternative. Claude models are also sold through Amazon Bedrock and Google Cloud Vertex AI, where usage may count against cloud commitments you already hold, depending on how those agreements define eligible spend. Our Bedrock versus direct calculator compares the two routes.
• Keep a termination right. A termination for convenience clause on the usage commitment makes the exit path real.

In a Claude Enterprise deal the seat price is the headline. The renewal cap, the overage rate and the data clauses decide what the term costs.

When should each step happen before renewal?

On a first Claude contract, run the same sequence against the pilot and agree the four clauses before it ends, while you still have a choice of vendor. Our GenAI contracting red lines white paper lists the terms to hold firm on.

What to do next

1. List the four clauses. Mark the renewal cap, overage rate, model substitution and data use terms as open or closed in your current draft.
2. Cap the renewal. Fix a named percentage ceiling on the renewal uplift, covering seats, usage rates and discount.
3. Price the overage. Set the overage rate and add a true up right into the committed tier, with a notice threshold.
4. Tier the models. Tie the commercial rate to a capability tier and add successor pricing for replacement models.
5. Write the data terms. Put training, retention and deletion protections into the contract itself.
6. Size from the pilot. Base the committed volume on observed pilot usage.
7. Benchmark first. Price a credible alternative before you sign.

Frequently asked questions