OpenCode on Amazon Bedrock: open weight coding agents

The post arrives as AWS argues the constraints of mainstream coding assistants have become a real friction point: sending proprietary code to a third-party API, locking into one model provider, or paying per-seat regardless of usage. OpenCode, built in Go, reads and edits files, runs shell commands, and understands project structure through Language Server Protocol diagnostics, connecting to over 75 LLM providers including Bedrock.

The pitch rests on a division of labor between models rather than a single best one. In the sample configuration, planning and architecture tasks route to Kimi K3, which reasons on every turn and supports a 1M-token context, while code generation goes to Nemotron 3 Super 120B, whose Mixture-of-Experts design activates only 12B of 120B parameters per token. GPT-OSS 120B serves as a default for multi-file implementations. The article points readers to the Artificial Analysis Coding Index and Bedrock Evaluations to compare models on their own prompts, and cites CrowdStrike's fine-tuned NVIDIA Nemotron reaching 96% valid query accuracy against GPT-4o at 61% and Claude Sonnet 4.5 at 94%. Ethara.AI, an AWS customer, runs this in production with an orchestration layer that routes requests by capability. The open question is whether cost and residency settings line up: Flex tier promises 50 percent lower cost for batch work, but the cheaper global profile and strict data residency may pull in opposite directions for regulated teams.