AI Inference Pricing: AWS, Azure and Google Cloud Costs

Published Token Rates Require a Defined Workload

Prices were checked on 15 November 2026 and are expressed in USD. These are selected published examples, not equivalent-model benchmarks or an exhaustive model catalogue. The advantage and buyer columns are editorial assessments of the billing arrangements.

Do not substitute another provider’s direct API rate for a missing Azure number. Microsoft says actual prices can vary by agreement, purchase date and exchange rate, and directs customers to its calculator or sales team. Microsoft says actual Azure prices can vary by agreement, purchase date and exchange rate, and directs customers to its calculator or sales team. Microsoft Foundry pricing.

For procurement, request the exact model version, deployment geography and billing currency together. Apply the same discipline used when comparing Microsoft 365 vs Google Workspace pricing: define the purchased service before comparing its headline price. A Global deployment and a regional deployment should remain separate purchasing options until their operational requirements match.

Batch and Caching Change Different Parts of the Bill

Batch discounts reward workloads that can wait. AWS limits its advertised discount to selected foundation models; Azure ties its offer to supported Batch API deployments. Neither statement establishes a blanket discount for every model, region or interactive request. Check eligibility before putting a discount into an approved budget. AWS pricing, Azure pricing.

Caching instead targets reused input. AWS documents implicit caching as best effort: repeating an identical prompt does not guarantee a hit. Explicit caching requires model-specific controls and qualifying prompt prefixes. Consequently, a forecast should distinguish observed cache reads from content merely intended for reuse. AWS caching documentation.

Azure’s current documentation adds a material qualification to Azure AI Foundry costs. Models before GPT-5.6 do not charge extra for cache writes; GPT-5.6 and later families can incur write charges alongside discounted reads. Microsoft states that charges for 12-hour prompt-cache writes apply from 1 November 2026. A purchasing worksheet that records only cached-read savings can therefore omit a relevant meter. Microsoft caching documentation.

Google distinguishes implicit caching from explicit caching. Creating a cache generally incurs standard input-token charges unless a model-specific cache-write rule applies; explicit caching also incurs storage charges based on retention time. Implicit caching has no separate storage charge. Google caching documentation.

For Gemini 3.5 Flash, Google lists global cached input at $0.15 per million tokens against $1.50 for standard input. Explicit caching also incurs storage charges based on retention; the current price page should be checked for the chosen model and region. Google pricing.

Treat cache retention as an architectural purchasing decision. Alongside business password manager pricing, assess who controls retained business information and whether the proposed reuse pattern is operationally acceptable.

Provisioned Capacity Makes Utilisation a Purchasing Variable

Provisioned inference changes the comparison from consumed tokens to purchased capacity. AWS describes Provisioned Throughput as fixed-cost capacity purchased in model units and directs buyers to their account manager for unit specifications and pricing. Its documentation also says billing continues until the provisioned resource is deleted, subject to commitment restrictions. AWS throughput documentation.

Azure charges provisioned deployments by deployed provisioned throughput units per hour. It offers hourly billing and reservations, prorates partial hours and states that deployments cannot be paused: deployment billing stops when they are deleted. Evaluate reservation obligations separately before treating deletion as the end of every financial commitment. Microsoft throughput billing.

Google describes Provisioned Throughput as a fixed-cost, fixed-term subscription. Buyers specify the supported model and available location. Its overview identifies critical, consistently high-throughput workloads as candidates, but does not disclose a numerical subscription rate on that page. Google throughput documentation.

Request a capacity proposal using measured traffic, prompt lengths and output requirements. Compare its committed cost with the pay-as-you-go cost of the same workload, then test how idle periods affect the result. Keep security expenditure visible alongside inference, using the separate endpoint security pricing comparison when preparing the broader technology budget.

Choose by Workload, Geography and Measured Cost

Choose an AWS Bedrock pilot if your shortlist includes its hosted models and the quoted AWS region meets your requirements. Choose Azure OpenAI if an Azure deployment is required, but obtain the missing numerical rates before approving expenditure. Test Google’s Standard and cached-input arrangements if repeated context is central to your application.

For document processing that can wait, evaluate eligible batch options. For customer-facing applications, benchmark response time and task quality before considering a capacity commitment. These are procurement recommendations, not evidence that any provider is universally faster or cheaper.

Require each proposal to identify model version, region, input and output meters, cache reads and writes, storage, deployment mode and commitment terms. Award the contract on measured cost per accepted business result within the required operating conditions. A low token price becomes commercially useful only when the application produces acceptable work at that price.