Amazon Bedrock now exposes a set of open‑weight large language models that can be called directly from the OpenCode terminal‑native AI coding agent. Engineers can keep source files, prompts, and model responses inside their own AWS account, avoid dedicated GPU infrastructure, and eliminate per‑seat licensing fees.
Open‑Weight Model Support in Bedrock
Bedrock provides managed endpoints for models such as Moonshot AI Kimi K3, OpenAI GPT‑OSS 120B, and NVIDIA Nemotron 3 Super 120B. Selecting a model is a single API parameter change, allowing a single OpenCode binary to route different coding tasks to the model that best fits the workload without redeploying code.
Integrating OpenCode with Bedrock
OpenCode runs locally on a developer workstation or CI runner, reads and edits files, executes shell commands, and uses LSP diagnostics to understand project structure. When it needs a language model, it invokes the Bedrock SDK over the AWS network. The request inherits the caller’s IAM policy, is recorded in CloudTrail, and can travel through PrivateLink for VPC‑isolated traffic. All payloads are encrypted in transit and at rest per Bedrock defaults.
Cost and Compliance Considerations
Agentic coding workflows can multiply token usage by five to thirty times, making token‑price a dominant cost factor. Open‑weight models on Bedrock are priced per‑token, so consumption scales directly with actual usage, avoiding flat per‑seat charges. Bedrock also offers region‑specific inference profiles: the global profile (global.moonshotai.kimi-k3) routes requests to any supported AWS region, while a geographic profile such as us.moonshotai.kimi-k3 confines processing to the United States. These profiles support data‑residency requirements and align with compliance programs including HIPAA, SOC 2, ISO 27001, FedRAMP, and GDPR.
Operational Implications
Because the models are serverless, there is no need to provision or patch GPU instances. Scaling is handled by Bedrock, but practitioners must still manage IAM permissions that allow OpenCode to call bedrock:InvokeModel. CloudTrail logs provide an audit trail for each inference request, useful for post‑incident analysis or cost attribution. Multi‑model workflows—e.g., using a smaller model for linting and a larger model for code generation—can be orchestrated by configuring OpenCode to select the appropriate endpoint per task. Ethara.AI’s production deployment demonstrates that such orchestration can be extended to multiple agents for research‑grade pipelines.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
- Adopt OpenCode as a thin client; the heavy lifting stays in Bedrock, reducing local resource requirements.
- Define IAM policies that grant only the necessary
bedrock:InvokeModelactions and scope them to the required model ARNs. - Choose an inference profile that matches your data‑residency and cost goals; prefer the global profile for lower cross‑region fees when compliance permits.
- Instrument CloudTrail and, if needed, VPC flow logs to monitor usage patterns and detect unexpected token spikes.
- Plan multi‑model task routing to balance performance and cost, leveraging the ability to switch models via a single API parameter.
