Four platforms, four ways of metering the same thing. Cost per task, commitment sizing, model routing, data terms, and the eleven moves that decide the number.
How to Negotiate with OpenAI and Anthropic: The Vendors With Nobody to Call
Fewer than 50 sales reps globally per vendor, focused on $100M+ deals. Below $10M a negotiation rarely starts, discounts run 5 to 25 percent on commitment size, and the only leverage is credible competition between OpenAI, Anthropic, and Gemini with a benchmarked case.
What You Are Actually Buying
Part 1 of the Negotiating Anthropic series. Tokens, seats and three routes to purchase, each priced differently. The model tier choice that moves cost more than any discount, and why input and output are not the same commodity.
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 buy answers, summaries and decisions. Everything difficult about these contracts follows from that gap.
Four platforms carry most enterprise demand, and each meters differently enough that a straight price comparison is meaningless.
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.
See the CIO playbook for negotiating OpenAI contracts, the AI Platform Contract Playbook, the Azure OpenAI versus direct OpenAI comparison, and the Vendor Shield always on program.
Five things decide the outcome, and they apply across all four vendors.
See the multi vendor negotiation scorecard and the benchmarking practice.
OpenAI sells two things that get confused in procurement. ChatGPT Enterprise is a per seat subscription for people. The API is metered per token for systems. They have different economics and should be negotiated separately.
On the seat side, the question is how many people genuinely use it weekly, not how many were provisioned. Adoption on enterprise rollouts is uneven, and paying for dormant seats is the most common waste we find.
On the API side, input and output tokens are priced differently, and output usually costs several times more. That single fact should shape how prompts and responses are designed, and it rarely does.
Committed spend buys a better rate and costs you flexibility. Ask for price protection, ask what happens if a cheaper model supersedes the one you committed against, and get the answer in writing. See the CIO playbook for negotiating OpenAI contracts.
Anthropic prices per token across a model range, and the gap between the tiers is wide. Routing everything to the most capable model is the single most expensive habit in enterprise AI, and it is also the easiest to fix.
Two mechanics change the effective rate more than any discount will. Prompt caching cuts the cost of repeated context, which matters enormously for retrieval heavy applications. Batch processing trades latency for a lower rate on work that does not need an immediate answer.
Model these before you negotiate. A workload restructured around caching and batching can land below the rate you were trying to argue for, without committing to anything.
Anthropic also sells seats for Claude Enterprise alongside API access. Treat those as two negotiations, as with OpenAI. See the GenAI vendors advisory practice.
Google's AI pricing rarely stands on its own. Vertex AI usage tends to be folded into the wider Google Cloud commitment, which is an advantage or a trap depending on how the commitment is written.
The advantage is that AI spend can count toward a commitment you were making anyway. The trap is that it becomes invisible, buried inside a cloud number nobody itemizes at renewal.
Insist on separate reporting for AI consumption even when it sits inside a broader commitment. You cannot negotiate a line you cannot see, and a year of unattributed spend is a year of lost leverage.
Check how AI spend interacts with your committed use discounts and whether it counts at full value. See the Google Cloud advisory practice.
Bedrock's proposition is access to several model providers through one API and one bill. For a buyer running a multi model strategy, that convenience is genuinely valuable, and it is also a form of lock in worth pricing.
Two consumption modes matter. On demand charges per token with no commitment. Provisioned throughput reserves dedicated capacity for a fixed period, which lowers the effective rate and commits you to a specific model.
Provisioned throughput is where buyers get caught. Committing capacity to a model that is superseded three months later leaves you paying for something you no longer want to use.
Negotiate the right to move committed capacity between models on the platform. Bedrock's whole pitch is model choice, so a commitment that removes that choice deserves an argument. See the AWS advisory practice.
On regulated workloads these terms carry more risk than the price does, and on roughly half the deals we reviewed nobody had read them until security asked late in the process.
Three questions cover most of the exposure. Where is the data processed and stored. How long is it retained, and can you set that to zero. Is any of it used to train or improve models, and is the opt out contractual rather than a setting in a console.
A configuration toggle is not a contractual commitment. Settings change, defaults change, and a vendor is not bound by a checkbox. Get the answer written into the agreement.
Residency answers differ by vendor and by region, and they change as new regions come online. Ask for the current position in writing at signature and again at renewal. See the AI Platform Contract Playbook.
These are the moves we run on an enterprise AI negotiation. The first three do most of the work.
See the GenAI vendors advisory.
The eleven moves, cost per task normalization across OpenAI, Anthropic, Google and AWS Bedrock, commitment sizing, and the data terms that belong in the contract rather than a settings page.
Used across more than five hundred enterprise clients. Independent. Buyer side. Built for teams sizing an AI commitment against consumption they can actually evidence.
Every vendor quoted us a token price and every price looked different. Redress made them all quote the same workload end to end, and the ranking changed completely. We also found we were sending trivial work to the most expensive model available.
Material saving against the opening quote, and a routing policy we should have written a year earlier.
We work for the buyer. Always. There is no other side of our table.
OpenAI commercial signals, Anthropic commercial signals, Google AI commercial signals, AWS Bedrock commercial signals, and the broader enterprise AI commercial leverage signals across the practice.
The common advice is to pick one vendor, commit to volume, and negotiate the lowest token rate. We disagree. In the commitments we reviewed, single vendor lock removed the ability to route each task to the cheapest model that could do it, which mattered far more than the headline rate. The largest savings came from tiering workloads across providers and reserving the top model for tasks that needed it. The buyer side move is to normalize quotes to cost per task, keep at least two providers viable, and commit only on the proven baseline. Chasing one low rate through full commitment, the instinct vendors encourage, is how budgets lock to capacity a changing roadmap cannot use.
Source: Redress Compliance advisory engagement file, 2024 to 2025.
Enterprise AI vendor pricing models at a glance
| Vendor | Primary metric | Commitment lever |
|---|---|---|
| OpenAI | Per seat plus per token API | Enterprise committed use |
| Anthropic | Per seat plus per token API | Committed spend tiers |
| Google Vertex and Gemini | Per token plus per seat | Committed use discounts |
| AWS Bedrock | Per token by model | Provisioned throughput |
The major enterprise AI vendors price through a mix of per seat subscriptions for their apps and usage based per token rates for API and cloud platform access. OpenAI, Anthropic, Google, and AWS all separate human seat licensing from machine token consumption. Most enterprises run a blended bill of seats plus platform tokens.
Staying multi vendor preserves negotiating leverage and lets you route each workload to the best priced model, but it adds integration and governance overhead. A single vendor simplifies operations yet weakens your price position at renewal. Most large buyers keep at least two viable providers live to retain leverage.
Compare on cost per outcome, normalizing token rates, context limits, and seat fees against the same representative workload. Headline per million token prices hide differences in context window and model tiering. Run one real use case across each vendor before trusting list price comparisons.
Data residency, training data use, indemnification, and price protection are the terms that matter most in an enterprise AI contract. Confirm in writing that your prompts and outputs are not used to train shared models. These clauses, not the unit price, are where the largest long term risk sits.
Negotiate once usage is predictable enough to commit a meaningful annual volume, usually after a pilot phase of a few months. Committed spend unlocks better unit pricing across all four major vendors. Avoid signing a large multi year commitment before real consumption data exists, because AI usage patterns shift quickly.
