--- title: "AI Cost Math for Hospitals: Per-Seat vs Usage-Based in 2026" slug: "ai-cost-math-for-hospitals-per-seat-vs-usage" author: "ibl.ai Engineering" date: "2026-05-30 10:00:00" category: "Premium" topics: "AI cost healthcare, hospital AI pricing, HIPAA AI, prior authorization AI, clinical documentation AI, per-seat vs usage-based, ChatGPT Enterprise healthcare, GPT-5 pricing, Claude Opus pricing, self-hosted healthcare AI" summary: "What AI actually costs a hospital in 2026 — token pricing across the latest models (Claude Opus 4.7, GPT-5, Gemini 3 Pro, Llama 4), per-seat SaaS math, and why $60-per-clinician scales the wrong way for prior auth and clinical documentation." banner: "" thumbnail: "" --- ## Per-Seat Pricing Was Built for Software You Use Occasionally A mid-size health system has 5,000 clinicians. ChatGPT Enterprise lists at around $60 per user per month. That's **$300,000 per month** — $3.6M per year — before a single prior-authorization letter is drafted. The pricing model was built for collaboration software (Slack, Notion, Salesforce) — tools where most seats sit idle most of the day and the per-seat fee approximates "access." For AI that actually does work — drafting prior auths, summarizing visit notes, triaging messages — the seat model breaks. The cost scales with how many people *could* use it, not what they *do*. The same workload, priced by tokens consumed, costs a fraction. The math is the post. ## What the Latest Models Actually Cost in 2026 Token pricing across the major providers, approximate as of mid-2026 (always check provider docs for current rates):

Model	Provider	Input ($/MTok)	Output ($/MTok)	HIPAA-eligible?
Claude Opus 4.7	Anthropic	$15	$75	Yes (BAA)
Claude Sonnet 4.6	Anthropic	$3	$15	Yes (BAA)
Claude Haiku 4.5	Anthropic	$1	$5	Yes (BAA)
GPT-5	OpenAI	$10	$30	Yes (Enterprise BAA)
Gemini 3 Pro	Google	$3.50	$10.50	Yes (Vertex BAA)
Llama 4 (70B, self-hosted)	Meta (open weights)	~$0	~$0	Yes (you control PHI)
DeepSeek-R1 (self-hosted)	DeepSeek (open weights)	~$0	~$0	Yes (you control PHI)

For self-hosted open-weight models, "~$0 per token" means the marginal cost is just the GPU time. A single A100 or H100 instance ($1–3/hour reserved) handles thousands of clinical requests per day. ## A Real Workload: Prior Authorization at 5,000-Clinician Health System Prior authorization is the highest-volume, highest-pain administrative AI use case in any health system. A mid-size system processes roughly **10,000 prior-auth requests per month**. Each request is about 500 tokens in (patient context, clinical justification) and 1,500 tokens out (drafted letter with citations to medical-necessity criteria). For a deeper per-letter cost breakdown — including per-transaction specialty vendors (Cohere Health / Olive / Notable) and three scale tiers (community / regional / IDN) — see **[What AI Prior Authorization Actually Costs in 2026](/blog/what-ai-prior-authorization-actually-costs-2026)**. That's **5M input + 15M output tokens per month** for the entire prior-auth workload — across 5,000 clinicians, that's an average of 2 requests per clinician per month, with heavy concentration on a few high-volume specialties. ### What it costs by deployment shape

Deployment	Pricing shape	Monthly cost	Annual	PHI residency
ChatGPT Enterprise	Per-seat ($60/user)	$300,000	$3,600,000	OpenAI cloud (BAA)
Microsoft 365 Copilot	Per-seat ($30/user)	$150,000	$1,800,000	Microsoft cloud (BAA)
Direct API — Claude Sonnet 4.6	Token-based	~$240	~$2,880	Anthropic cloud (BAA)
Direct API — GPT-5	Token-based	~$500	~$6,000	OpenAI cloud (BAA)
ibl.ai self-hosted (Llama 4 / DeepSeek-R1)	Flat license + GPU	~$3,000–5,000	~$36,000–60,000	Inside your VPC / on-prem

The ibl.ai row covers the GPU instance, the platform license, and ongoing support. It does **not** include the BAA conversation, the vendor risk review, or the re-architecture every time the vendor updates their data-processing terms — because there is no third-party vendor in the data path. The model runs on infrastructure you already own. ## Why the Per-Seat Math Doesn't Work in Healthcare Three reasons per-seat AI fails harder in healthcare than anywhere else: **1. Usage is concentrated.** A handful of high-volume specialties (oncology, cardiology, GI) generate most of the prior-auth and documentation load. Buying a seat for every clinician means subsidizing the ones who barely touch it for the ones who hit it constantly. Token pricing aligns the bill to the actual work. **2. The clinical workforce is large and lower-paid than the cost model assumes.** A 5,000-clinician system isn't 5,000 attending physicians — it's nurses, techs, residents, schedulers, coders, billers. The seat fee assumes a uniform "knowledge worker" who can absorb $60/month of overhead. For a coding clerk doing prior auth all day, $60/month is fine; for a triage nurse who touches AI twice a week, it's not. **3. PHI residency forces re-purchase, not extension.** When a managed AI vendor updates its data-processing terms — or when the FDA / OCR publishes new guidance — every BAA gets re-papered. With self-hosted, the data never leaves; the model swap is a config change, not a procurement event. ## What Stays the Same, What Changes Self-hosting the runtime doesn't mean rebuilding the platform. The chat UI, the clinician dashboards, the audit logs, the model-routing-with-fallbacks, the multi-agent orchestration — all of that stays managed by ibl.ai. The compute, the model, and the PHI move inside the hospital's perimeter. The trade-off most health systems don't realize: **the per-seat SaaS line item is bigger than the all-in self-hosted infrastructure budget.** A $3M/year ChatGPT Enterprise contract pays for an internal AI platform team, dedicated GPUs, and the model-choice flexibility that comes with owning the stack — with money left over. ## Run the Numbers for Your Health System For workload-specific calculations — prior auth, clinical documentation, patient messaging triage — use the **[AI Help Desk Cost Savings Calculator](/resources/calculators/ai-help-desk-savings-calculator)** as a starting point (the math generalizes to most high-volume clinical-administrative workloads). For the deployment comparison side-by-side — including HIPAA posture, BAA reach, and air-gapped options — see **[Self-Hosted AI vs ChatGPT Enterprise for Healthcare](/resources/comparisons/self-hosted-ai-vs-chatgpt-enterprise-for-healthcare)**. For the full HIPAA-aligned architecture (Managed VPC → on-premise → air-gapped tiers, Epic / Cerner / athenahealth integrations, TCO at 10K clinicians), read **[Healthcare AI Reference Architecture on ibl.ai](/blog/healthcare-ai-reference-architecture)**. ## Why Family-Owned and New York Matters Here The sovereignty argument falls apart if the vendor on the other side of the BAA is on a five-year exit clock, foreign-owned, or acquired before the next OCR audit. ibl.ai is family-owned and operated from New York, NY — a long-term partner for U.S. health systems, defense, and regulated buyers, with a perpetual platform license and no investor exit pressure. The runtime is open source. The data stays inside the covered boundary. The math works at 100 clinicians or 50,000.