Enterprise AI pricing in 2026 is a moving target across four distinct metrics: per-token API pricing, per-seat productivity (Copilot, ChatGPT Enterprise, Claude for Enterprise), committed-use compute (Bedrock, Vertex, Azure OpenAI), and outcome-based contracts at the frontier. The variance across vendors at the same workload is wider than most procurement teams realise — and the renewal posture differs substantially. This is the buyer-side benchmark across 340+ engagements.
Per-token API pricing is the metric that anchors most enterprise AI procurement. The pricing model is consistent across vendors — input tokens and output tokens, priced per million — but the actual rates vary by an order of magnitude across model tiers, and the rate-to-capability ratio shifts every quarter as new model releases ship. As of mid-2026, frontier-tier models (GPT-5, Claude Opus 4.6, Gemini 2.5 Pro) price input tokens in the $8-15/M range and output tokens in the $30-75/M range; mid-tier models (Sonnet, Gemini Flash, GPT-4o-class) price 5-10x lower; small models (Haiku, Gemini Flash Lite) price another 5-10x lower.
The rate that procurement teams quote in benchmarking exercises is typically the input rate — but output tokens dominate cost for most enterprise workloads. The ratio runs 3-5x output-to-input on a typical RAG implementation, 8-12x on agent workloads, and 15-30x on code-generation use cases. A 5% discount on input rate combined with a 0% discount on output rate is functionally a 1% discount on actual spend. The negotiable position is to anchor benchmarking on a representative workload mix, not on the published price card.
Most enterprise API contracts above $500K annual commitment carry committed-use discount tiers — typically 10-25% off list at $1M annual commitment, 25-40% off at $5M, and material additional discounts above $10M. The structure differs between vendors: OpenAI and Anthropic offer enterprise discounts directly with no third-party intermediation; Microsoft Azure OpenAI bundles the discount inside the Azure MACC; AWS Bedrock includes Bedrock spend in EDP commitments. The vehicle choice changes the effective price, which is why we run AI commitments through structured software contract negotiation rather than accepting the published committed-use tier as the ceiling.
The pricing variance across vendors at the same workload is wider than most procurement teams realise.
The per-seat productivity AI market has consolidated around three primary vendors at the enterprise tier: Microsoft Copilot (formerly M365 Copilot, now multi-product), OpenAI ChatGPT Enterprise, and Anthropic Claude for Enterprise. Google's Workspace AI sits inside the Workspace bundle. Per-seat pricing list is $30/user/month at Microsoft Copilot and broadly comparable rates at the competing platforms, with enterprise discount typically arriving at the 5,000+ seat tier.
Microsoft Copilot pricing is the most complex per-seat tier in the AI market because the same name covers materially different products: M365 Copilot (the productivity-app AI), Copilot for Sales/Service (Dynamics-specific), Copilot Studio (the agent-building platform), Copilot for GitHub Enterprise, and the Copilot+ device-class licensing. Each prices separately and bundles with different base SKUs. Customers benchmarking Copilot need to scope the actual SKU mix before comparing prices.
OpenAI ChatGPT Enterprise prices broadly in line with Microsoft Copilot per seat, with the enterprise-tier deal structure including custom token quotas and admin/audit controls. Anthropic Claude for Enterprise is priced similarly with the Anthropic enterprise tier offering custom context windows, data-handling commitments, and seat-tier discounts at scale. The functional differentiation across the three has narrowed in 2025-2026; the procurement decision is now driven more by integration footprint and data-governance position than by capability gaps.
The hyperscaler-bundled AI offerings (AWS Bedrock, Google Vertex AI, Azure OpenAI Service) resell frontier models from third-party providers with the hyperscaler's cloud governance, compliance and committed-use overlay. For Anthropic models, all three resell. For OpenAI models, Azure OpenAI is the primary path. For Google models, Vertex is direct. The per-token list pricing across hyperscalers is broadly similar — Bedrock, Vertex and Azure OpenAI typically publish identical or near-identical rates for the same model — but the actual delivered price differs because the discount vehicle differs.
AWS Bedrock spend counts toward AWS Enterprise Discount Programme (EDP) commitments. For customers with active EDP, this is the lowest-friction way to consume frontier AI inside the AWS commercial framework — the spend draws down against the EDP commitment and inherits the EDP discount. For customers without an EDP, Bedrock prices roughly at list, with limited stand-alone negotiation flexibility.
Azure OpenAI Service spend counts toward the Microsoft Azure Consumption Commitment (MACC). Microsoft's posture is materially more flexible than AWS's on standalone Azure OpenAI commitments — multi-year deals direct with Azure OpenAI are common — but the MACC overlay is the most discount-efficient vehicle for customers running broad Azure footprints.
Google Vertex AI spend counts toward Google Cloud committed-use discounts. Google's posture on AI committed-use is the most aggressive of the three hyperscalers, with multi-year Vertex commitments routinely discounting 30-45% off list at $1M+ annual commitment.
The AI vendor market is experimenting with outcome-based pricing structures that price against business outcome (resolved support tickets, automated invoice processing, code-review acceptance) rather than per-token consumption. The structures are early — most enterprise AI procurement still anchors on the consumption metrics — but the outcome-based path is appearing on RFPs at increasing rates through 2025-2026. The pricing implications are wide: outcome-based contracts shift the consumption risk from buyer to vendor, but typically at a premium per-outcome rate that requires careful unit-economics modelling.
A specific variant emerging in 2026 is the per-agent-task metric — pricing AI workloads at the granularity of a completed agent task rather than the underlying token consumption. The metric works well for clearly-bounded workflows (resolved customer support tickets, processed expense claims) and poorly for open-ended workflows (research agents, code-generation agents). Customers procuring AI agents in 2026 should expect the per-agent-task metric to be on the table; the negotiable position is whether the task definition is narrow enough to be auditable.
The 2026 playbook covering AI contract red flags, data rights, IP ownership and usage-based pricing.
AI vendor pricing comparison is hard because the underlying products differ in capability, latency, throughput, context window and integration footprint. A naive comparison of $/M tokens at list price misses 80% of the actual delivered cost variation. The correct benchmark is a representative workload run across each vendor's actual offering, with the per-vendor discount overlay applied, and the integration cost (custom data connectors, vendor-specific SDKs, retention policy differences) included in the TCO.
We benchmark OpenAI, Anthropic, Google, Microsoft Copilot, AWS Bedrock and the hyperscaler bundles across 340+ engagements.
Our AI & SaaS Procurement practice covers OpenAI, Anthropic, Google, Microsoft Copilot, AWS Bedrock and the hyperscaler bundles.
Weekly compliance intelligence for IT leaders.