Our take
Amazon Bedrock is the enterprise-grade hyperscaler route to hosted models, with strong compliance credentials and a broad catalogue spanning first-party, downloadable, and third-party proprietary options. It suits regulated workloads that need verified data controls and AWS-native integration.
Use this for regulated workloads requiring independent security attestation, zero-retention, and EU data residency. Pick it if your organisation is already in AWS and wants unified billing and IAM integration. Skip it if you need predictable throughput on large downloadable models, or if your budget cannot absorb the steep price jump from lightweight first-party to large open-weight or frontier third-party options.
- Strongest compliance and data-control stack in sample: independent security attestation, zero-retention option, no training on prompts, and EU-region endpoint all verified rather than claimed.
- Competitive first-party pricing for lightweight models — Amazon Nova Micro 1.0 is roughly half the input price of Nova Lite 1.0 and cheaper than Claude 3 Haiku and GPT-5 6 Luna on this provider.
- Broadest model access among hyperscalers in this sample: 46 tracked offers spanning Amazon first-party, OpenAI, Meta, Anthropic, and Qwen families.
- Fastest measured throughput on a small downloadable model: OpenAI GPT-OSS Safeguard 20B at 232 transactions per second, 1.74× the second reading on the same model and 2.53× the reading for Amazon Nova Lite 1.0.
- Throughput collapses on large downloadable models: Meta Llama 3.1 70B Instruct at 20 transactions per second is 11.6× slower than OpenAI GPT-OSS 120B and 11.6× slower than the fastest GPT-OSS Safeguard 20B reading.
- Premium pricing on large downloadable and frontier third-party models: Meta Llama 3.1 70B Instruct is 20.6× the input price of Amazon Nova Micro 1.0 and 4.8× the input price of GPT-OSS Safeguard 20B; Claude 3 Haiku output is 8.9× Nova Micro output.
- Inconsistent throughput on identical model configurations: spreads range from 3.0% on Nova Micro 1.0 to 114.8% on GPT-OSS Safeguard 20B.
Privacy & data handling
- ✓
- Yes
- ✗
- No
- No answer on record
The colour says whether the answer favours you, not whether it is a yes — not training on your prompts earns a green cross, no zero-retention option earns a red one.
Prompt retention: none.