Amazon Bedrock vs Google Cloud Vertex AI: $0.035 Nova Micro vs $2.74/hr GSU Commitments

At the lightweight tier, Amazon Bedrock holds a decisive price advantage. Amazon Nova Micro costs just $0.035 per 1M input tokens and $0.14 per 1M output tokens, making it over 7x cheaper on input than Vertex AI's entry-level Gemini 3.1 Flash-Lite ($0.25 input / $1.50 output). For mid-tier production workloads, Bedrock's Nova Pro ($0.80 / $3.20) and Vertex AI's Gemini 3 Flash ($0.50 / $3.00) trade blows closely.

At the frontier reasoning tier, Vertex AI undercuts Bedrock's third-party flagship offerings. While Anthropic Claude 3.5 Sonnet ($3.00 / $15.00) and Claude Opus ($15.00 / $75.00) command premium rates on Bedrock, Vertex AI prices Gemini 3.1 Pro at $2.00 input and $12.00 output per 1M tokens. However, finance teams must watch Vertex AI's 200K context cliff: any prompt exceeding 200,000 tokens immediately doubles the input rate to $4.00/1M and raises output to $18.00/1M.

Buying dedicated throughput reveals a stark architectural divide. Amazon Bedrock sells Model Units (MUs) priced around $21.00 to $50.00+ per hour depending on the model and commitment term (No-Commit, 1-month, or 6-month). Because a single MU often translates to $15,000–$36,000+ per month, teams must maintain 60% to 70% continuous 24/7 utilization just to break even against On-Demand rates. Worse, exceeding your MU quota triggers hard throttling exceptions unless engineers manually build fallback routing to On-Demand APIs.

Google Cloud Vertex AI abstracts reserved capacity into Generative AI Scale Units (GSUs), starting at just $2.74 per GSU-hour on a 1-year commitment (with 1-month and 3-month terms available). Each GSU provides a fixed universal throughput budget that burns down at published model-specific ratios. Crucially, Vertex AI features native Pay-As-You-Go spillover: when traffic bursts past your reserved GSU ceiling, excess tokens automatically bill at standard On-Demand rates with zero dropped requests.

For asynchronous workloads—such as nightly document summarization, synthetic data generation, or bulk classification—both Amazon Bedrock and Vertex AI offer a flat 50% discount on Batch Inference. Because neither platform enforces a minimum monthly spend or high token floor for batch jobs, the financial break-even volume against On-Demand is immediate at 1 job (a single JSONL or BigQuery input submission), provided your application can tolerate up to a 24-hour completion window.

For real-time global traffic, both cloud providers have eliminated the historical TCO penalty of multi-region failover. Bedrock's Cross-Region Inference and Vertex AI's global endpoints dynamically route burst requests across available regional data centers with zero cross-region network egress markup, billing strictly at the source region's standard token rate. This allows engineering teams to absorb regional quota spikes on On-Demand APIs without over-provisioning expensive dedicated Model Units or GSUs.