Enterprise AI licensing in 2026. What OpenAI, Anthropic, Google and AWS charge for.

Compare enterprise AI vendors on cost per completed task, since token rates are not comparable across platforms. Most savings come from routing, caching, batching and contract terms, and little from the headline discount.

Enterprise AI is the only major software category where the unit of pricing does not match the unit of value. Vendors sell tokens and seats. You are buying answers, summaries and decisions, and most of what makes these contracts hard follows from that gap.

Four platforms carry most enterprise demand: OpenAI, Anthropic, Google and AWS Bedrock. Each meters differently enough that setting their price lists side by side tells you little about what a workload will cost.

How do OpenAI, Anthropic, Google and AWS charge for enterprise AI?

All four sell two separate things: seats for people and metered tokens for systems. Seats are a per user subscription. Tokens are billed per million, with input and output priced separately and output usually costing several times more than input.

Bought direct, an OpenAI or Anthropic commitment is a contract with the model maker. Google and AWS sell models inside a cloud relationship, where AI spend lands on a commitment you already hold. For the Microsoft route to OpenAI models, see our Azure OpenAI versus direct OpenAI comparison.

What decides the final price of an enterprise AI contract?

Five things decide the outcome, and they apply to all four vendors. The rate card feeds only the first of them.

1. The unit of comparison. Cost per completed task, normalized across vendors, rather than cost per million tokens.
2. The commitment. How much you promise to spend, measured against a baseline you can evidence from real usage.
3. The routing. Which model tier handles which workload. This is where most of the money sits.
4. The data terms. Retention, training use and residency, which carry as much risk as price on regulated work.
5. The exit. What happens when a model is deprecated, or a better option appears halfway through the term.

Our multi vendor negotiation scorecard rates a proposed deal on each of these, and the benchmarking practice tests the terms against what comparable buyers have signed.

How do you compare AI vendor quotes fairly?

Price one representative task end to end on each platform, then divide by the share of attempts that produce a usable result. Token rates mislead because models spend different numbers of tokens on the same job, and reasoning models bill their thinking as output.

A worked example: 100,000 clause reviews a month

Say your legal team checks 100,000 contract clauses a month against its standard positions. Each call sends 6,000 input tokens: 5,000 of fixed instructions that never change, plus a 1,000 token clause extract. The models and rates are hypothetical, set to resemble a top tier, a middle tier and a low priced reasoning model.

Monthly cost is the cost per attempt times 100,000, divided by the usable rate, because failed attempts are rerun. The cached rows leave out the small charge for writing the cache. Model C has the lowest token rates and still costs more than twice as much per task as Model B, since it spends 4,500 output tokens reasoning and answering.

Check context pricing too: Gemini 2.5 Pro on Vertex AI charges more for prompts over 200,000 tokens.

Caching discounts also differ. OpenAI bills cache hits at a tenth of the input rate on its current models and Anthropic at a tenth or less, while Vertex AI takes 75 percent off cached Gemini 2.5 tokens and adds an hourly storage charge.

Why we would not open with a demand for a lower token rate

The usual advice is to press for the deepest per token discount and accept a larger multi year commitment to get it. We think that gets the order wrong. In the example above, caching and batching cut Model A's monthly cost by 74 percent, a larger cut than rate discounts usually deliver.

List prices also fall often enough that a long rate lock can leave you above the market by the second year. Restructure the workload first, commit on what the restructured workload consumes, and spend the negotiation on price protection and exit terms.

What should you negotiate with each AI vendor?

With OpenAI and Anthropic, the main issue is the split between seats and API. With Google and AWS, it is how AI spend sits inside the wider cloud commitment.

OpenAI: ChatGPT Enterprise seats and API tokens

ChatGPT Enterprise is a per seat subscription for people, and the API is metered per token for systems. Procurement teams often merge them, but they have different economics and different waste, so negotiate them separately.

• Seats. Count how many people use ChatGPT each week and compare it with the number provisioned. Ask for the right to reduce seats at each anniversary.
• API. Output tokens usually cost several times more than input. That should shape how prompts and answers are designed, and it rarely does.
• Commitment. Committed spend buys a better rate and costs you flexibility. Get in writing what happens if a cheaper model supersedes the one you committed against.

Our CIO playbook for negotiating OpenAI contracts covers the clauses.

Anthropic: model tiers, caching and batch

Anthropic prices per token across a model range, and the gap between tiers is wide. Routing everything to the most capable model is the most expensive habit in enterprise AI, and the easiest to fix.

Two mechanics change the effective rate more than any discount will. Prompt caching bills repeated context at a fraction of the input rate, which matters most for retrieval heavy applications. The Batch API halves prices for work that can wait, and Anthropic documents that the two discounts stack.

A workload restructured around caching and batching can land below the rate you were trying to argue for, without committing to anything.

Claude Enterprise seats and API access are two negotiations, as with OpenAI. Data terms also carry a price: on Claude 4.6 and later models, restricting inference to the United States adds 10 percent to every token category. Our GenAI vendors advisory practice handles both sides.

Google: Vertex AI inside the Google Cloud commitment

Google's AI pricing rarely stands on its own. Vertex AI usage is usually folded into the wider Google Cloud commitment, which helps or hurts depending on how that commitment is written.

• The advantage. AI spend can count toward a commitment you were making anyway.
• The trap. AI spend disappears into a cloud total that no one itemizes at renewal, and you lose the evidence to negotiate it.
• The check. Confirm how AI spend interacts with your committed use discounts, and whether it counts at full value.

Insist on separate reporting for AI consumption even inside a broader commitment, because you cannot negotiate a line you cannot see. Vertex AI also sells Provisioned Throughput, which raises the same model lock in question as Bedrock. See our Google Cloud advisory practice.

AWS Bedrock: model choice and Provisioned Throughput

Bedrock offers several model providers through one API and one bill. That convenience has real value for a company running more than one model, and it is also a form of lock in worth pricing.

• On demand. Charges per token with no commitment.
• Provisioned Throughput. Reserves dedicated capacity for one model, billed hourly, with no commitment or a 1 month or 6 month term. Longer terms lower the hourly price.

Committing capacity to a model that is superseded three months later leaves you paying for something you no longer want to use. Bedrock's pitch is model choice, so ask for the right to move committed capacity between models. See our AWS advisory practice and Bedrock pricing guide.

What should an enterprise AI agreement say besides the price?

It should fix the data terms, protect the price for the full term, and say what happens when a model is retired. These clauses are cheaper to win before signature than at any point after it.

Which data terms belong in the contract?

On regulated workloads, retention, training use and processing location carry more risk than the price does. Three questions cover most of the exposure.

1. Where is the data processed and stored?
2. How long is it retained, and can you set retention to zero?
3. Is any of it used to train or improve models, and is the opt out in the contract or only a setting in a console?

The published defaults have improved. OpenAI states that API data has not been used for training by default since March 1, 2023, keeps abuse monitoring logs for up to 30 days, and offers zero data retention to customers it approves.

A configuration toggle is still not a contractual commitment. Settings change, defaults change, and a vendor is not bound by a checkbox. Residency options also vary by vendor and region, so ask for the current position in writing at signature and again at renewal.

Contract wording to ask for

• Price protection with pass through. Your rate holds for the term, and any cut to the vendor's list price reaches you automatically.
• Deprecation cover. Notice longer than the public policy, and the right to move committed spend to the successor model. Anthropic publishes at least 60 days' notice for publicly released models, far shorter than a multi year commitment.
• Portable commitment. Committed spend usable across models, and ideally across seats and API, so a routing change does not strand money.
• Rate limits at the committed level. Write the peak throughput you need into the order.
• Indemnification. Cover for third party intellectual property claims arising from outputs, with the exclusions read by counsel.
• Separate AI reporting. Monthly consumption by model and project whenever AI spend sits inside a cloud commitment.

The AI Platform Contract Playbook sets out the wording, and our note on OpenAI data privacy clauses covers that vendor line by line.

What will the AI vendor's account team say, and how should you answer?

Expect pressure to commit more, provision more seats and rely on published defaults. The replies below keep each conversation tied to measured usage and written terms.

What have we seen in recent enterprise AI negotiations?

Done properly, this work takes 20 to 35 percent out of enterprise AI spend. Almost none of that comes from arguing about the headline rate. It comes from how workloads are routed and how the contract is structured.

• Dormant seats. Paying for seats that are provisioned and not used weekly is the most common waste we find.
• One model for everything. Sending every request to the most capable tier is the most expensive habit we see, and tiering is usually the fastest saving.
• Unread data terms. In roughly half the deals we reviewed, no one had read the data terms until security asked late in the process.

Which steps cut enterprise AI spend the most?

These are the 11 steps we run on an enterprise AI negotiation. The first three do most of the work, and none needs the vendor to agree to anything.

1. Normalize every quote to cost per task. Price a representative workload end to end with each vendor.
2. Tier your workloads before you negotiate. Decide which tasks need the top model. In most companies it is a minority.
3. Model caching and batching first. Restructuring the workload often beats anything you will win at the table.
4. Separate seats from API. They are different purchases with different waste.
5. Audit seat activity, not seat count. Weekly active use is the number that matters.
6. Commit only on the proven baseline. Never on the roadmap, which changes faster than the contract does.
7. Keep two providers viable. A working integration you could route traffic to next quarter, so the alternative is real.
8. Get price protection in writing. Include what happens when the vendor's own prices fall, which they regularly do.
9. Cover model deprecation. Establish what you are entitled to when the committed model is retired.
10. Negotiate rate limits alongside price. A cheap rate you cannot consume at peak is worth little.
11. Put data terms in the contract. Retention, training use and residency belong in the agreement, and a settings page does not count.

How do you check what your company actually uses?

Every platform reports usage by model, so most of the evidence is already in these consoles.

• OpenAI. The Usage and Costs views in the API platform dashboard by project and model, and user analytics in the ChatGPT Enterprise admin console.
• Anthropic. Usage and cost reports in the Claude Console by workspace and model.
• Google. The Cloud Billing export to BigQuery, filtered to Vertex AI SKUs and grouped by project label.
• AWS. Cost Explorer filtered to Bedrock, and the CloudWatch InputTokenCount and OutputTokenCount metrics per model.

If AI spend is growing inside a cloud bill with no attribution, fix the tagging before renewal talks start. The software spend health check then sizes AI spend against the rest of your software portfolio.

How does Redress help with enterprise AI contracts?

We work only for the buyer. Our Enterprise AI scoping is a six week engagement that maps consumption, normalizes vendor quotes to cost per task and sets the commercial plan; contact us to start one.

• Negotiation. We run the commercial conversation across OpenAI, Anthropic, Google and AWS Bedrock, including the data terms.
• OpenAI advisory. Focused work on seats, API commitments and price protection.
• Vendor Shield. Vendor Shield, our always on advisory subscription, covers the AI platforms and the wider cloud contracts.

What to do next

1. This month. Export three to six months of usage from each platform and tag it by workload.
2. Before any quote. Price one representative task end to end on at least two vendors, including the rerun rate.
3. Before you commit. Test caching, batching and a cheaper tier on your largest workloads, and size the commitment on the result.
4. On seats. Set the renewal number from weekly active users, not provisioned seats.
5. Early in the negotiation. Send the data terms to security and counsel, with retention, training use and residency on the redline list.
6. At signature. Check that price protection, deprecation cover, portable commitment and rate limits are in the signed order.

Frequently asked questions