Amazon Bedrock

The wiki's first cloud reseller / aggregator — the provider category the domain names (Azure/Bedrock) but had no page for. Bedrock is not a model maker: it's a fully managed AWS service that resells hundreds of foundation models from many providers behind one API, the "middle layer" between the labs and the application. It is the managed shape of the ai-gateway; the other is the self-hosted router you run yourself (omniroute).

The reseller model

• Multi-provider catalog: Anthropic Claude, Meta Llama, Mistral, Cohere, and Amazon's own Nova / Titan — i.e. it aggregates the very labs this wiki pages individually, plus first-party models.
• One API, serverless, pay-as-you-go — no infrastructure to manage; switch or evaluate models "as new models emerge" without re-plumbing, the explicit anti-lock-in pitch.
• Scale: powering generative AI for 100,000+ organizations.

Why it matters to the market thesis

Bedrock is where the synthesis's "compatibility as strategy" and "engineering sets real cost" threads land structurally: the reseller commoditizes model choice into a config parameter, turning the labs into interchangeable backends and pushing differentiation up to routing, caching, and data integration (exactly the cost levers the synthesis flagged). It also reframes the open-vs-closed axis: through a reseller, a closed Claude and an open Llama are the same kind of API call — the buyer experiences the whole price/openness spectrum through one pane. It's the demand-side mirror of the supply-side provider map: the labs (openai, deepseek, qwen, mistral-ai, xai-grok, google) make models; resellers like Bedrock (and Azure AI Foundry) are how most enterprises actually consume them.

Re-checked 2026-10-03 (homepage)

The 100,000+ organizations figure still stands. Two things are new on the page since July:

• "Amazon Bedrock Managed Agents, powered by OpenAI": OpenAI's frontier models and its agent harness, run on AWS infrastructure. The section above lists openai among the labs a reseller sits beside; the homepage now sells OpenAI through Bedrock. The page does not say which OpenAI models are included or on what terms, so this page records the offer, not its scope.
• AgentCore, an agent platform for "any framework and model", with tracing and evaluation built in.

The homepage also quotes its own figures for Guardrails (blocks "up to 88%" of harmful content), distillation ("up to 500% faster", "up to 75%" cheaper) and prompt routing ("up to 30%" savings). These are vendor claims with no method on the page; they are recorded here as claims, not results.

The shift fits the thesis above, and pushes it one step further: the reseller now resells the harness as well as the model, so the differentiation that moved up to "routing, caching, and data integration" is moving into the agent layer too.

Related

llm-provider · llm-api-pricing · llama · mistral-ai · openai · deepseek