Comparing cloud AI model pricing in 2026 means comparing per-token rates across Amazon Bedrock and Google Cloud’s managed AI platforms, where the spread between the cheapest and most expensive models now exceeds a factor of forty on the same platform. The cheapest open-weight text models on Bedrock’s on-demand tier, such as DeepSeek v3.2, are billed at $0.62 per 1M input tokens and $1.85 per 1M output tokens, while flagship-tier models like Claude 3.5 Sonnet list at $6.00 input and $30.00 output — a ratio that makes model selection, not vendor selection, the dominant cost lever for most engineering budgets.

This article is a working cost comparison for practitioners building on managed model APIs. If you are still framing the consumption decision itself, start with our practical engineering guide to what cloud AI is; the analysis below assumes you have chosen a hosted inference tier and need to know what it will cost. The baseline concepts — managed APIs, tuning platforms and raw accelerated infrastructure — map directly onto the pricing structures compared here, and the platform-selection criteria in our breakdown of how to choose your cloud AI platform set the frame for the vendor-level differences that follow.

What one million tokens costs

Token pricing is the unit economics layer of every cloud AI deployment, and the published rates vary by model class, region and throughput tier. The table below consolidates on-demand rates from current vendor pricing pages, using the US-region standard tier where regional splits exist.

Two structural facts stand out. First, on a single vendor’s marketplace, the DeepSeek v3.2 on-demand rate of $0.62 per 1M input tokens sits roughly ten times below the Claude 3.5 Sonnet input rate of $6.00, and roughly 48 times below its output rate — which is why routing rules that send easy traffic to cheap models are the highest-ROI optimization available before any architecture change. Second, when you drop below the managed API layer to self-hosted serving on GPU infrastructure, billing flips from tokens to compute time, and the comparison becomes a utilization problem rather than a rate-card problem.

Batch, caching and discounts

The rate card is only the starting price. Amazon Bedrock prices batch inference for select foundation models at a 50% lower price compared to on-demand inference pricing, which means any workload tolerating minutes or hours of latency — enrichment jobs, embeddings regeneration, evaluation suites — should be scheduled as batch by default. The same platform’s Claude 3.5 Sonnet v2 listing shows prompt-cache economics explicitly: $7.50 per 1M tokens for cache writes and $0.60 per 1M tokens for cache reads against the $6.00 on-demand input baseline, so retrieval-heavy applications with repeated system prompts can cut effective input cost dramatically by designing for cache hits rather than re-sending context. Regional variation is real but bounded: the same DeepSeek v3.2 model is $0.62 per 1M input tokens in US regions and $0.74 in the batch of regions that includes Europe (Stockholm), a roughly 19% premium that matters for EU-residency-constrained deployments. Even the tooling has a meter: Bedrock’s simple prompt optimizer runs at $0.03 per 1,000 tokens, charged on the input prompt plus the resulting optimized prompt.

For Google Cloud, the discount structure attaches to commitment rather than batching: Agent Platform training and prediction carry no minimum usage duration and are charged in 30-second increments, and consuming GPU capacity from a reservation spreads billing across two SKUs so committed-use discounts apply to AI workloads the same way they apply to ordinary compute.

Model catalog and residency

Price comparisons are only actionable if the model you selected is deployable in the region your compliance posture requires. Microsoft’s Foundry documentation confirms that models sold by Azure are hosted and operated by Azure, and that the catalog now spans Azure OpenAI models plus selected models from providers like Cohere, DeepSeek, Meta, Mistral AI and xAI — with availability varying by region and cloud. The practical consequence for EU teams is that the multi-provider marketplaces (Bedrock and Foundry) have become the shortest path to mixing cheap open-weight models with frontier models under a single data-processing agreement and a single residency envelope. Bedrock’s own Claude 3.5 Sonnet extended-access listing already includes Europe (Frankfurt), Europe (Ireland), Europe (Zurich) and Europe (Paris) among its supported regions, so the premium open-model tiers are no longer a US-only procurement.

A cost-control procedure

The teams that keep cloud AI bills predictable follow roughly the same sequence:

1. Classify workloads by latency tolerance. Everything that can wait goes to batch inference at the discounted rate; only interactive traffic pays on-demand prices.
2. Set a routing policy before setting budgets. Cheap open-weight models handle classification, extraction and first-pass drafting; frontier models verify, escalate and produce final copy.
3. Design prompts for cache hits. Long stable prefixes — system instructions, tool definitions, retrieved context — should exploit cache-read pricing instead of re-paying the full input rate on every request.
4. Instrument per-team token tagging from week one. Inference is the line item that scales quietly; per-workspace spend controls and usage analytics are now standard even on vendor business plans, and enterprises can negotiate volume discounts on top.
5. Re-run the rate-card comparison quarterly. Model tiers, service tiers (standard, flex, priority, reserved) and regional availability shift fast enough that a comparison older than a quarter is stale.

The bottom line for practitioners

Cloud AI pricing has converged on a pattern every infrastructure engineer will recognize: a cheap commodity tier, an expensive premium tier, and a set of commitment and caching levers between them. The decisions that actually move spend are which tier each workload lands on, whether traffic is routed to batch and cache where possible, and whether the serving region matches both the price and the residency requirement. Teams that treat the model catalog as a portfolio to be routed, rather than a single vendor bet to be placed, will find the 2026 rate cards work in their favor; teams that pick one flagship model and send all traffic to it will pay the premium on every request.

Sources

• Amazon Bedrock pricing — AWS
• Gemini Enterprise Agent Platform pricing — Google Cloud
• Foundry Models sold by Azure — Microsoft Learn
• Business Pricing — OpenAI