Commit Less, Keep the Exit: Negotiating the Enterprise AI Platform Contract

OpenAI doubled GPT-5.5 list pricing to $5 input and $30 output per million tokens in April 2026. Any AI commit priced on 2025 economics is already wrong.

Prepared by Redress Compliance  ·  June 2026  ·  Representative enterprise AI estate scenario (benchmark scenario, not a quote)

Enterprise software pricing normally moves a few percent a year, upward, on a renewal calendar. AI platform pricing moves in both directions, by integer multiples, between your signature and your first true up.

The GPT-5.5 release doubled list rates overnight for on demand API customers. Meanwhile cached input pricing fell to $0.50 per million tokens and batch processing held at half of standard rates.

Both directions hurt an unprotected contract: a price increase reprices your uncommitted consumption immediately, because on demand rates float with the published list by default. A price decrease strands your prepaid commit above market, because the commit was sized in dollars against yesterday's rates. The contract, not the forecast, has to absorb this.

Three clauses convert volatility from your problem into the provider's problem:

• A per token rate lock for the full term, covering the contracted model class rather than a named model.
• A most favored pricing clause on successor models: when a new generation ships, your rate per unit of benchmarked capability cannot worsen.
• A downward repricing trigger: if published list falls below your contracted rate, your rate follows within a defined window.

Providers resist the third clause hardest, which tells you where they expect prices to go. A vendor confident that rates will rise sells rate locks cheaply. Price the silence accordingly.

Four platforms can credibly anchor an enterprise AI estate today. They differ less on model quality, which converges every quarter, and more on commercial structure: how you commit, what draws down, and who carries repricing risk.

The structural read: OpenAI sells scarcity, Anthropic sells routing economics, Google sells bundling, AWS sells consolidation. Your leverage with each is different.

With OpenAI you negotiate repricing protection; with Anthropic, effective rates through caching and batch defaults. With Google you negotiate inside the Workspace and GCP envelopes you already own; with AWS, commit eligibility and model unit flexibility.

Every enterprise tier now states that your prompts and outputs are not used to train foundation models by default. That sentence is necessary and insufficient. The negotiated carve out has to reach three places the default language often does not.

• Subprocessors and affiliates. The training exclusion must bind every subprocessor and affiliate in the chain, including any cloud marketplace intermediary. A carve out that binds only the contracting entity leaks at the first subcontract.
• Derived artifacts. Fine tuned model weights, embeddings, evaluation sets, and telemetry derived from your usage are your confidential information. The contract should say so explicitly and survive termination, with deletion certified.
• Retention windows. Standard API terms typically retain request logs for around 30 days for abuse monitoring. Regulated buyers should negotiate the zero data retention endorsement where offered, or a contractual cap with audit rights where it is not.

The zero data retention endorsement is a deal desk checkbox, not a product limitation. In our file it goes unrequested far more often than it gets refused.

Output ownership is the easy clause: every major provider now assigns output rights to the customer on paid tiers. The hard clauses are the conditions stapled to the indemnity behind it.

All four providers offer some form of copyright indemnity for output on enterprise terms. Each indemnity carries conditions that can quietly empty it. The recurring carve outs:

• Outputs the customer modified after generation.
• Outputs generated with safety filters disabled.
• Outputs where the customer knew of likely infringement.
• Content supplied through retrieval from the customer's own corpus.

Treat the indemnity as an engineering requirement, not just a legal one. If the indemnity is conditioned on filters being on, your platform team must log filter state per request, or the condition is unprovable when a claim arrives. In our file, fewer than one in four clients could evidence filter state before we raised it.

Model rights deserve equal attention: your right is to a model class and a capability level, not a binary. Contract for: named model access at signature, successor model access at no worse unit economics, and a defined notice period before any model you depend on is deprecated. Section 7 prices the deprecation half of this.

The provider's opening structure is predictable: a multi year prepaid commitment, sized on an enthusiastic adoption forecast, discounted just enough to feel bought. The math rarely survives contact with measured consumption.

Across the engagement file, 2024 to 2025, first proposal commitments ran 2.1x to 3.4x the client's measured run rate, and 62 percent of prepaid commitments reached anniversary with unused balance. Unused balance is not savings at a discount; it is prepayment for consumption that never existed.

Benchmark ranges: Redress Compliance advisory engagement file, 2024 to 2025.

The worked scenario below models a representative estate: a 12,000 employee financial services enterprise running 3,000 assistant seats at $60 per seat per month, or $2.16M per year.

The API workload adds 20 billion input and 4 billion output tokens per month. At GPT-5.5 list that is $220,000 per month, $2.64M per year, for a combined list run rate of $4.8M per year.

Worked representative estate. Benchmark scenario, not a quote. Benchmark ranges: Redress Compliance advisory engagement file, 2024 to 2025.

Four commit mechanics decide whether the middle and bottom paths are achievable:

• Drawdown scope: every product (seats, API, fine tuning, embeddings) should draw from one commit pool.
• Rollover: unused balance is forfeited at anniversary by default. A 12 month rollover of 25 to 30 percent of balance is regularly granted in our file when asked before signature.
• True up cadence: quarterly, not annual, so growth is paid at the negotiated rate rather than re anchored at list.
• Ratchet floors: strike any clause that lets the commit only move upward at renewal.

Where we disagree with the standard advice. Resellers and account teams preach the same line: maximize the discount by signing the longest prepaid commitment you can. We disagree.

Price per useful token has been falling 30 to 60 percent per year through new model generations, caching, and batch routing. A 25 percent discount on a three year prepay is a poor trade against that curve. The buyer side move is the shorter term with rate protection, sized to measured run rate, growth captured at quarterly true up.

Multi provider architecture is usually justified as resilience: model outages, deprecations, and capability regressions stop being existential. The commercial case is at least as strong. Discount outcomes in our file track the number of credible alternatives at the table more tightly than any other variable, including deal size.

Each credible alternative at the table is worth roughly ten points of reduction. Credible is the operative word; a logo on a slide moves nothing. A second provider becomes leverage when a real workload runs on it, your prompts and evaluations are portable, and your platform team can shift traffic inside a quarter.

The cost of credibility is real but modest: an abstraction layer and duplicated evaluation suites typically add 5 to 8 percent engineering overhead in the estates we have benchmarked. Against a 20 point swing on a multimillion dollar commit, the arithmetic is not close. Bedrock and Vertex make this cheaper still, since one contract carries several model families.

Enterprise AI SLAs cluster around 99.9 percent uptime with service credits as the sole remedy. Credits are not the point; they refund pennies against an outage that idles a production workflow. Negotiate instead for termination rights on chronic failure (for example, three missed months in any six) and commit relief proportional to downtime.

Model deprecation is the more expensive failure mode, and the one a generic SLA never covers. Providers retire models on roughly 6 to 12 month notice cycles, and a deprecation forces regression testing, prompt rework, and sometimes architecture change on your side at your cost.

Without that clause, a deprecation mid term quietly converts your rate lock into a repricing event. The replacement model arrives at new list, and your protection expired with the old one.

The provider playbook across all four vendors is consistent enough to table. None of these moves are improper; all of them are priced for the unprepared.

Sequence matters as much as the responses. The leverage points above are nearly all preserved or lost before commercial negotiation formally begins, which is why we run the process on a 120 day calendar.