Amazon Bedrock Azure AI Foundry and Vertex AI Pricing and Catalog Comparison

Amazon Bedrock, Microsoft’s Azure AI Foundry, and Google’s Vertex AI have settled into distinct identities as of September 16, 2026. While all three provide on-demand access to frontier AI models through cloud APIs, they diverge on catalog size, pricing, and specific model performance. This comparison of Amazon Bedrock, Azure AI Foundry, and Vertex AI shows that the platforms now represent different bets on enterprise infrastructure.

Amazon Bedrock emphasizes breadth, offering more than 100 models from 18-plus providers. Azure AI Foundry leans on catalog scale, citing more than 10,000 models with approximately 50 new additions every month. Vertex AI maintains a more curated Model Garden of 200-plus models, focusing on deep integration for Google’s Gemini family.

The pricing gap between flagship models is currently the widest point of divergence. Amazon Bedrock’s flagship reasoning model, Claude Opus 4.6, is priced at $5.00 per 1 million input tokens and $25.00 per 1 million output tokens. Azure AI Foundry’s GPT-6 Astra carries a higher rate of $10.00 for short-context input and $50.00 for output, rising to $20.00 and $75.00 respectively for long-context workloads. Google’s Gemini 3.8 Flash, which reached general availability in 2026, is priced at $0.75 for input and $3.75 for output per 1 million tokens.

A daily workload processing 10 million input tokens and 2 million output tokens illustrates the practical cost differences. On Claude Opus 4.6 through Bedrock, this volume costs $100 a day. The same volume on GPT-6 Astra’s short-context tier costs $200 a day. Running that workload through Gemini 3.8 Flash reduces the cost to about $15 a day. This pricing difference is cited as a primary driver for FinOps teams to route high-volume, latency-tolerant tasks to Flash-tier models while reserving flagship reasoning models for complex requests.

Performance benchmarks from August and September 2026 show varied strengths across the platforms. Anthropic’s Claude Opus 5 holds a 97.00% score on the SWE-bench Verified tracker, though Bedrock listings still reference Claude Opus 4.6 as the flagship Opus-tier offering. GPT-6 Astra leads on agent-oriented benchmarks, scoring 69% on AutomationBench-AA and 59% on Terminal-Bench v4.0. Gemini 3.8 Flash trails on raw SWE-bench at 80.0%, but maintains a 90.2% score on MMLU-Pro and 94.4% on GPQA Diamond.

The platforms also differ in how they count their available models. The 10,000-plus figure for Azure AI Foundry includes fine-tuned variants, small open-weight checkpoints, and third-party models from providers like NVIDIA. Bedrock’s catalog of 100-plus models is a more curated list where each model has its own dedicated model card and pricing row. Vertex AI’s Model Garden includes the Gemini and Gemma families alongside third-party options like Claude and Llama.

Operational commitments vary by service. Amazon Bedrock and Azure AI Foundry both offer a 99.9% monthly uptime SLA for their invocation APIs. Vertex AI provides a 99.5% SLA for the Gemini Enterprise Agent Platform online inference, though several other Vertex AI services carry a 99.9% guarantee.

Each platform has also developed native agent frameworks. Bedrock offers Agents and AgentCore, which now includes generally available web search grounding. Azure AI Foundry provides the Foundry Agent Service for multi-agent orchestration. Vertex AI utilizes the Gemini Enterprise Agent Platform. Capacity management also differs, with Bedrock using provisioned throughput per model unit and Azure offering Provisioned Throughput Units (PTUs) that are portable across different models.