OpenAI contract negotiation for CIOs in 2026. Size the commitment low and write the protections in.

OpenAI rewards buyers who commit conservatively and keep a credible second model. Size the commitment on optimized realized usage, tie the rate to published list prices, and put data, indemnity and deprecation terms in the order form.

OpenAI contracts are short, priced in units that change every few months, and sold by an account team that expects you to commit before you know your usage. This guide is for the CIO, procurement lead or AI platform owner about to sign or renew an OpenAI agreement, whether for ChatGPT Enterprise seats, API consumption or both.

Our GenAI vendor practice negotiates these agreements for enterprise buyers and takes no money from OpenAI or Microsoft.

How does OpenAI enterprise pricing work in 2026?

OpenAI sells on three commercial streams. ChatGPT Enterprise is per seat licenses plus a shared credit pool, the API is billed per token consumed, and a negotiated spend commitment usually sits across both.

Published API pricing is quoted per million input and output tokens, by model, and it changes often. Any committed rate you sign should name that published page, because it is the benchmark your price reset clause will point at.

• API. Billed per input and output token, by model, with batch, flex and priority service tiers at different rates. OpenAI renamed priority processing to Fast mode on July 30, 2026, and requests tagged priority now run in Fast mode.
• ChatGPT Enterprise. A per seat fee for core features, plus a shared credit pool for advanced features, on an annual commitment.
• Commitment. A spend floor that buys a discount and a lock, which you then have to manage across both streams.

What do ChatGPT Enterprise seats and credits include?

Under OpenAI's flexible pricing model, a seat covers the core experience: chat, search, file upload and canvas. Advanced features draw from a credit pool bought at the contract level.

Deep Research, thinking models, image generation, advanced voice and Codex all consume credits. On the Enterprise plan every user draws from one shared pool, and there is no default per seat cap. Credit expiration is set contract by contract, which makes expiry and rollover terms you can negotiate.

What happens when the credit pool runs out?

When the pool runs dry, what happens depends on three settings. Decide who owns them before the first invoice arrives.

• Pause. Advanced features stop until more credits are added.
• Overages. The workspace owner enables overages, and an overage limit of zero blocks credit requests once the committed pool is used up.
• Top up. More credits are bought through the account team, at whatever rate your order form sets.

Usage alerts and overage limits sit in billing settings, where the workspace owner can also set spend controls by group.

Why price the three streams separately?

Ask for seats, credits and API consumption as separate lines on the order form. A single blended number hides which stream is overcommitted, and in the contracts we reviewed it was almost always the credit pool or the API floor, rarely the seats.

What have we seen in recent OpenAI and GenAI contract negotiations?

Between 2024 and 2025 we advised on roughly 20 to 30 enterprise OpenAI and generative AI contracts. Pricing moved fast, and buyers often signed commitments before anyone understood their real usage. The same three problems appeared in most of them.

• Overcommitment. First year committed spend was set 30 to 60 percent above realized usage, with no flexibility on drawdown.
• Locked token prices. Rates were fixed at signature while list prices fell 20 to 50 percent over the term, and no clause brought the committed rate down with them.
• Boilerplate accepted. Data, security and indemnity terms were signed as standard, although 6 of 10 of those buyers could have negotiated them.

How do you build a token cost model before you commit?

Price your real workload against the published per token rates, model by model, before you accept any commitment number. In July 2026 list input rates spanned a 150x range, from $30.00 on the premium reasoning tier to $0.20 on the smallest routing model. That spread matters far more than the discount percentage.

Three modifiers apply to those list rates. Batch runs 50 percent below list, cached input bills at 10 percent of the input rate, and priority processing (now Fast mode) runs at roughly twice the standard rate.

How fast did those list prices move?

They moved within weeks. On July 30, 2026 OpenAI cut gpt-5.6-terra by 20 percent to $2.00 input and $12.00 output, and gpt-5.6-luna by 80 percent to $0.20 input and $1.20 output. By late September the pricing page led with the GPT-6 family, including gpt-6-sol at $2.00 and $10.00.

A buyer who fixed a Terra rate in early July was paying 25 percent above list by August. Your cost model should therefore carry two numbers per model: today's list rate and a planning assumption for where list will sit at mid term.

What does a worked example look like?

Take an internal assistant that processes 1 billion input tokens and 200 million output tokens a month. On gpt-5.4 that is $2,500 of input and $3,000 of output, or $5,500 a month at list. The table shows what engineering choices do to that bill before any discount.

Routing the workload to gpt-5.4-mini cuts the bill 70 percent with no negotiation at all. Caching at an 80 percent hit rate takes the gpt-5.4 input line from $2,500 to about $700: 800 million cached tokens at $0.25 plus 200 million uncached tokens at $2.50.

Routing policy and caching change cost far more than any discount OpenAI will grant. Do not size a spend commitment until engineering has fixed the routing and caching design, or the commitment will be priced against the unoptimized number.

What happens if you commit before routing is settled?

Say you commit $66,000 a year, the first row of the table, because that was the pilot run rate. Engineering then routes the workload to gpt-5.4-mini before go live. Annual spend lands near $19,800, and $46,200 of the commitment has nothing to draw against unless the contract allows rollover or reallocation.

Gaps like this are why we set commitments against the optimized design and add an expansion option at the same rate. Our guide to GenAI consumption billing and token cost control covers the internal controls that keep the forecast honest once you are live.

Which two benchmarks should you check before committing?

Benchmark against your realized usage and against the published list price trajectory. The gap between commitment and usage and the gap between your rate and list both compound in OpenAI's favor. A commitment sized on forecast at a fixed rate overpays twice: once on volume you never use, and again on a unit price the market has already left behind.

Where do OpenAI enterprise contracts go wrong?

They go wrong in three places: overcommitted spend, token rates locked at signature, and data and indemnity boilerplate accepted without review. The table sets OpenAI's typical first draft against the position you should hold.

Each row is a separate negotiation with a deadline. Once the order form is signed, the deal desk treats all of them as settled, so run them in parallel instead of trading them away one at a time.

How should you set the spend commitment?

Floor it at conservative realized usage, measured from production traffic, and leave optimistic forecasts out of it. Overcommitment was the single most common loss in the contracts we reviewed. Pilot traffic is a poor guide, because pilots run on the largest model with no caching and a small, enthusiastic user group.

Why does the token price need a reset clause?

List rates keep falling, and a committed rate with no reset leaves you paying last year's price. Draft the reset so it works without goodwill: name the published pricing page, set a quarterly comparison date, and state that the billed rate is the lower of the committed rate or current list.

A reset that depends on a request, a review or vendor consent is one the account team controls. Our price hold clause language shows how to pair the reset with a cap on increases, so the clause protects you in both directions.

What happens to unused credits and seats?

Enterprise credit expiration is contract specific, and unused credits die at term end unless you negotiate rollover. Seats true up on growth but rarely true down, so a workforce reduction leaves you paying for empty licenses until renewal.

Write both into the order form. Credits roll forward into a renewal term, and seat counts can be reduced at renewal without repricing the seats that remain.

Which clauses decide the value of an OpenAI contract?

Four clauses carry most of the value: data usage, intellectual property indemnity, service levels and model deprecation protection. OpenAI's paper is short compared with an Oracle or SAP agreement, which tempts legal teams to wave it through. Whatever the document does not say defaults to vendor discretion.

The standard business terms also let OpenAI update the agreement and its policies by posting changes, with at least 30 days notice when a change materially affects your rights. That is the main reason to move the terms below into your order form, where they cannot change mid term.

Data usage and retention

OpenAI's enterprise data commitments are good defaults: no training on business data, customer ownership of inputs and outputs, and admin controlled retention in ChatGPT Enterprise. API inputs and outputs may be kept for up to 30 days for abuse monitoring.

Put these in the contract rather than relying on the default. The order form should cover four points.

• Training. OpenAI does not train on your inputs or outputs.
• Retention. Defined retention windows for each product you buy.
• Deletion. Deletion of your content on exit, with written confirmation.
• Policy changes. Later edits to OpenAI's privacy pages do not weaken these terms during the term.

API customers with regulated workloads should ask about zero data retention eligibility on qualifying endpoints and get the approved scope in writing. Our note on OpenAI data privacy clauses goes clause by clause.

Intellectual property indemnity

OpenAI's business terms include an indemnity for third party IP claims, the commitment OpenAI announced for output as Copyright Shield. The current terms exclude claims arising from your own content, from modifications or combinations OpenAI did not make, and from your own applications. Also ask how fine tuned models and knowingly infringing use are treated.

The web terms cap each party's liability at the fees paid in the prior 12 months, but carve indemnification out of that cap. Write the indemnity into the order form and keep it outside the general cap. If OpenAI insists on a limit, agree a dollar figure sized against your exposure.

Service levels and support

The standard paper carries thin uptime language for a system that may sit in your critical path. Ask for a defined uptime percentage with service credits, a named technical contact, and Fast mode (formerly priority processing) for latency sensitive workloads, which OpenAI prices at roughly 2 times the standard tier.

Model deprecation protection

OpenAI's published deprecation policy promises at least 6 months notice for generally available models, at least 3 months for specialized variants, and as little as 2 weeks for preview models.

Several retirements fall in 2026, each forcing migration work on the buyer.

• August 26, 2026. The Assistants API was discontinued, with the Responses and Conversations APIs as the migration path.
• October 23, 2026. The gpt-3.5-turbo, gpt-4 and o1 snapshots retire.
• December 11, 2026. The original gpt-5 and o3 snapshots shut down.

OpenAI can revise that policy, so write the notice floor into the agreement, add an equivalence commitment on price and capability for any replacement model, and require migration assistance when a retirement forces engineering work on your side. OpenAI names gpt-5.6-terra and gpt-5.6-sol as the replacements for the October retirements, at different prices.

• Data. No training, defined retention and deletion on exit, all in the order form.
• Indemnity. IP indemnity with exclusions you have read, kept outside the liability cap or under a cap you negotiated.
• SLA. A defined uptime number with service credits in place of best effort language.
• Deprecation. Contractual notice, price and capability equivalence, and migration assistance.

What contract wording should you ask OpenAI for?

Ask for specific wording on eight points, and send it in your first redline so it is on the table before price is discussed. The table lists each term and why it matters.

OpenAI's standard terms require notice of non renewal or scope reduction at least 30 days before the next renewal term, so put that date in the contract register on the day you sign. Our seven clauses to push back on gives redline wording for several of these.

Should you buy OpenAI direct or through Azure OpenAI?

Buy through Azure when you carry an undrawn Azure commitment or need data zone and regional deployment controls. Buy direct when you need the newest models first plus the ChatGPT Enterprise seat and credit stack. The same models ship in two commercial wrappers, and the wrapper changes your counterparty, your pricing units and your negotiating options.

How are Azure provisioned throughput units billed?

Azure's provisioned throughput units bill on deployed capacity by the hour, whatever you consume. Microsoft's PTU billing guidance states that deployments cannot be paused, billing stops only on deletion, and PTUs above the reservation bill at the full hourly rate.

Reservation scoping catches many Azure buyers. Reservations are bought separately per deployment type, so a global provisioned reservation cannot cover a data zone deployment, although one global reservation can consolidate PTUs across several regions. Create the deployment first and buy the reservation second, so you never pay for capacity Azure could not place.

Why does the MACC decide it for many CFOs?

Azure OpenAI consumption draws down a Microsoft Azure Consumption Commitment. An enterprise with an undrawn Azure commitment effectively pays for OpenAI models with money it has already promised to Microsoft. Our Azure MACC negotiation guide covers how to size that commitment with AI in it.

Even if you intend to buy direct, get a live Azure OpenAI quote for the same workload. Two commercial channels for the same models give you a comparison no single vendor negotiation can, and the Azure OpenAI and direct OpenAI comparison sets out the differences in detail.

What will the OpenAI account team say, and how should you answer?

Expect the same handful of lines in most OpenAI negotiations. Each has a factual answer, and having it ready keeps the conversation on your terms.

• "Commit more and we can improve the rate." Show the usage baseline after routing and caching, and offer to commit that number with an option to expand at the same unit rate.
• "We don't offer price reset clauses." Point out that pay as you go customers already follow the published page, where OpenAI's terms make price changes effective 14 days after posting. You are asking for your committed rate to follow the same page down.
• "Our data commitments are on the website, so they don't need to be in the order form." The business terms let OpenAI update its policies on notice. Ask for the current wording to be fixed for the term.
• "Credits expire at the end of the term. That is standard." OpenAI's own help center says credit allocation and expiration are defined in the order form, so rollover is a term you can write in. Ask for unused credits to carry into the renewal term.
• "The replacement model is better, so you won't need equivalence terms." Better on benchmarks does not mean the same price or the same results on your prompts. Ask for the replacement at an equal or lower rate plus a migration window.

How does a CIO build negotiating power with OpenAI?

Negotiating power comes from a credible multi model strategy and a usage forecast you can support with production data. Three elements matter most.

• Multi model. A live alternative on Anthropic or open models changes the discount you are offered.
• Phased commitment. Start small, expand on proven usage, and avoid a large year one floor.
• Benchmark clause. The right to reset to published pricing whenever it decreases.

Phasing matters as much as the alternative. A 12 month term with a mid term expansion option keeps you inside the falling price curve, while a three year lock at a fixed rate bets against the market trend. If you must sign multi year for a discount, do not trade away the price reset clause.

How does the approach change with company size?

A company with a few hundred ChatGPT Enterprise seats and modest API use has little room on unit price. Its value lies in credit rollover, seat reduction rights and a short term, so it can switch or resize after 12 months.

A company rolling ChatGPT Enterprise to 20,000 staff and running production API workloads has the volume to win rate concessions. It also carries more exposure to deprecation and to a stranded commitment, so it should spend its negotiating time on reallocation rights and the reset clause. Our Claude Enterprise contract clauses guide helps when pricing Anthropic as the live alternative.

Why we advise against committing big early to lock the discount

The usual advice is to commit big and early to lock the best generative AI discount before prices rise. We disagree, because prices have mostly gone the other way. In the contracts we advised on, the buyers who committed hardest paid the most per token by year two, since early committed rates stayed fixed while list rates fell.

The better course is to floor the commitment at conservative realized usage, insist on a clause that resets your rate to published pricing whenever it drops, and keep a live alternative model in production. In a falling market, a smaller commitment with a reset clause costs less over the term than a large lock.

When list prices are falling, the buyer with the largest commitment holds the weakest position at the table.

When should you start the OpenAI negotiation?

Start six months before signature or renewal. Each source of negotiating power takes weeks to build, and none of them can be created in the final two weeks.

Time the signature against OpenAI's quarter, not yours. Consumption vendors still run on sales quarters, and a deal that closes in the last two weeks of one wins concessions that the same deal in week three of the next quarter does not. Our note on quarter end timing for OpenAI and Anthropic deals covers the calendar.

Reveal your options in this order: the routing model mix first, the Azure quote second, the live alternative model last. Each carries more weight when OpenAI learns of it late in the process, and our GenAI practice runs them in exactly that sequence.

Questions to ask OpenAI before you sign

1. Which of our current models have a published or planned retirement date inside the term?
2. How is the committed rate adjusted if the published list rate falls?
3. Can unused commitment be drawn on credits, seats or any model, and does it roll into renewal?
4. What uptime figure and service credits apply to our tier, and do they cover Fast mode?
5. Which data retention settings apply to each product we are buying, and which endpoints qualify for zero data retention?

What to do next

1. Build the baseline. Forecast usage conservatively from realized production consumption, not pilots.
2. Fix the design first. Agree routing and caching with engineering before anyone discusses a commitment number.
3. Floor the commitment. Set it at the conservative number, with an expansion option at the same unit rate and rollover of anything unused.
4. Write the reset clause. Tie the committed token rate to OpenAI's published pricing page, checked every quarter.
5. Contract the data terms. Put no training, retention and deletion terms in the order form, in writing.
6. Protect against retirements. Add notice, price and capability equivalence, and migration assistance for model deprecation.
7. Price the alternatives. Request Azure OpenAI pricing for your workload and keep a live alternative model in production.
8. Calendar the renewal. Record the non renewal notice date and restart the usage baseline six months before renewal.

Frequently asked questions