When an enterprise evaluates ChatGPT Enterprise against Claude Enterprise, the number that actually decides the budget rarely appears in either vendor's marketing. Anthropic's published Enterprise price is US$20 per seat per month, billed annually with a 20-seat minimum, and then usage billed at API rates on top [2]. OpenAI, for its part, publishes a token-based rate card that applies only to eligible token-based Enterprise agreements, with the customer's own contract controlling the applicable rates and discounts [4]. The practical takeaway for procurement teams is that the offer on the table is a three-layer document: the commercial structure, the usage rate schedule, and the data-handling exceptions. A comparison that stops at the headline seat price misses two of the three layers, and the third layer is where most compliance exposure lives.
The commercial structure determines whether the seat price means anything
Anthropic's Enterprise plan is explicitly a two-part cost. The US$20/seat/month figure carries conditions that change its meaning: annual billing, a 20-seat floor, and usage cost that scales with the model and task [2]. The plan bundles SCIM provisioning, audit logs, a compliance API for observability and monitoring, custom data retention controls, network-level access control, IP allowlisting, and a HIPAA-ready offering, with no model training on customer content by default [2]. Anthropic's page also states plainly that prices and plans are subject to change at its discretion and that listed prices exclude applicable tax [2].
When comparing this against any other vendor, three questions belong in the evaluation sheet before any per-seat figure is compared: what is the minimum seat commitment, which rate card governs usage, and what usage is included versus metered separately. OpenAI's rate card confirms why this matters on the other side of the table: it covers chat models, ChatGPT Work and Codex models, and separately metered features such as voice, web search, and image generation, with a note that the rates apply only after a workspace's billing model changes and that the agreement itself prevails over listed discounts [4]. Two vendors with identical seat prices can produce very different invoices if one meters agent and voice usage by token and the other includes it.
Token rates span a wide range inside a single vendor, and multipliers move the total
Anthropic's current published API rates, in US dollars per million tokens, run from Fable 5.1 at $10 input and $50 output, through Opus 5.5 at $4 and $20 and Sonnet 5.5 at $2 and $10, down to Haiku 5.5 at $0.10 input and $0.50 output for prompts up to 100K tokens [2]. Within one vendor's lineup, Fable 5.1 and Haiku 5.5 differ by 100x on input. Batch processing saves 50%, and workloads that must run with US-only inference are priced at 1.1x for input and output tokens [2]. Prompt caching has its own read and write prices, reflecting a 5-minute TTL [2].
This is why a total-cost model should be built from the buyer's own traffic profile rather than from an output-only rate. A workload with a high cache hit rate and batch-friendly jobs will land far from the list price; a real-time, long-context workload may pick up multipliers. OpenAI's rate card shows the same structure on its side: long-context input above 272K tokens is billed at 2x the standard rate and long-context output at 1.5x, fast and ultrafast modes carry 2x to 6x standard rates, and regional data-residency processing adds a 1.1x multiplier [4].
The parallel worth noting is structural, not a claim that the two products are equivalent. Anthropic's 1.1x is a premium for US-only inference on its API [2]; OpenAI's 1.1x is a premium for regional data-residency processing [4]. Both vendors charge roughly a 10% premium for keeping inference within a specified geography, so a residency or locality requirement is not by itself a differentiator. Where the totals do diverge is speed and length: both charge meaningfully more for fast modes and long context, which means the workload mix must be estimated before signing. All of these figures are page-published rates subject to change; the defensible move is to lock the applicable rate schedule into the contract.
The data-retention promise is only as good as its exception list
Anthropic offers a zero-data-retention (ZDR) arrangement under which it does not store customer prompts or responses at rest after the API response is returned. The conditions matter as much as the promise: ZDR is enabled per organization, must be requested through the sales team, and does not automatically extend to other organizations under the same account [5].
The exception list is where an evaluator earns their salary. Batch processing carries 29-day retention because async storage is required; code-execution data can be retained up to 30 days; and even under a ZDR or HIPAA arrangement, Anthropic may retain inputs and outputs for up to 2 years where a chat or session is flagged by its automated trust and safety systems [5]. None of these exceptions makes ZDR worthless, but each one can defeat a specific compliance requirement if the buyer's workflow relies on batch inference or flagged-sensitive content. The request to the vendor should therefore be for the endpoint-by-endpoint coverage list and written confirmation of which exceptions apply, not a yes-or-no answer on ZDR.
Model selection and retention policy also interact. Claude Fable 5 and Claude Mythos 5 are designated Covered Models that require 30-day data retention, and ZDR is not available for either; an organization whose retention configuration does not meet the requirement gets a 400 invalid_request_error when calling the model [5]. A ZDR organization can enable 30-day retention for a specific workspace to use these models [5]. The procurement implication is concrete: if a business line needs true zero retention, the newest Covered Models are off the table for that line, and the model shortlist should be built per data-retention tier rather than per capability alone.
Verify the OpenAI entity from the DPA and the certifications from official channels
OpenAI's data processing addendum specifies the contracting entity by customer location: customers in the EEA or Switzerland contract with OpenAI Ireland Ltd, while customers in the UK or anywhere else contract with OpenAI, LLC [1]. This is a first-party, citable term that belongs directly in contract review, since it determines which entity holds the data-processing obligations and which jurisdiction's framework attaches.
Compliance claims are a different matter. Community forum responses assert that ChatGPT Enterprise and the API are SOC 2 compliant, with the API described as Type 2 [3][6]. These are forum and support-team paraphrases, not certifications, and they should not enter a procurement file as evidence. The correct artifact is the vendor's Trust Portal report and the latest audited SOC 2 report supplied under NDA. The same standard applies to the OpenAI side of the comparison that applies to Anthropic: a documented, first-party source or a written contractual commitment.
Turning the comparison into a checklist
Combining the layers, a defensible evaluation matrix for either vendor contains: the minimum seat commitment and billing term; the governing rate card and its effective date; batch, caching, speed-mode, and regional-processing multipliers that apply to your workload; the data-retention tier per model, including any Covered Model restrictions; the endpoint coverage and exceptions of any ZDR or HIPAA arrangement; the contracting entity by region; and compliance evidence sourced from official audit channels. Items that cannot be verified from primary documentation, such as OpenAI's ZDR scope or either vendor's GDPR transfer mechanisms, should be converted into written pre-signature questions rather than assumptions. Both vendors reserve the right to change prices, so the final step is to reconcile page prices against contract rates and file the contract as the source of truth. An offer that cannot answer every line of that matrix in writing is not yet ready for signature.