OpenAI on Amazon Bedrock

The practical consequence is straightforward: the primary technical barrier to multi-model evaluation is gone. Running a benchmark between Claude Opus 4.7 and GPT-5.5 no longer requires separate contracts, separate identity policies, or separate billing accounts. Bedrock's IAM-principal-level cost attribution means both models appear in the same CUR 2.0 reports under the same access controls. For teams that built their AI strategy around a single vendor because switching costs were too high, those switching costs just dropped materially.

The implications extend to agent orchestration. AWS AgentCore's managed agent harness — which spins up agents from a model, system prompt, and tools without orchestration code — is now accessible to OpenAI's models as well as Anthropic's. Teams that have already invested in AgentCore tooling don't need to rebuild to evaluate different frontier models. The orchestration layer stays; the model swaps.

Three things are worth acting on. First, audit AI procurement decisions made on the assumption of long-term vendor lock-in — particularly workloads split across clouds for reasons that no longer apply. Second, update your model evaluation process to take advantage of a unified control plane: benchmark the same task against multiple models using identical tooling, IAM policies, and logging pipelines. Third, review vendor contracts for exclusivity clauses negotiated when the market looked different — the Microsoft precedent suggests these arrangements are more renegotiable than they appeared six months ago.

The harder strategic question is whether your team treats model selection as a one-time architectural decision or an ongoing operational variable. The infrastructure now supports the latter. Organizing for it is the real work.