Does the output suffer when using other model providers with Claude Code harness?

Does the output suffer when using OpenCode Go models from Claude Code harness (through a proxy like LiteLLM) vs using them from OpenCode itself?

Yes, using OpenCode Go models (e.g., DeepSeek, GLM, Kimi) via a proxy in Claude Code will likely reduce answer quality and reliability compared to using them natively in OpenCode. [1 , 2 , 3 , 4 , 5 ]

While the raw model latency is comparable, the “agentic” performance degrades due to a mismatch between the model and the harness. [1 ]

Quality & Reliability (Significant Degradation) [1 ]

Harness-Model Mismatch: Claude Code is “vertically integrated” with Anthropic models. Its system prompts, tool definitions, and context compression logic are hyper-optimized for how Claude “thinks” and responds. [1 ]
Tool Calling Failures: When you force a different model (like GLM or DeepSeek) into the Claude Code harness, you often experience degraded tool-call reliability and “unexpected compaction failures.” The model may struggle to interpret Claude-specific instructions, leading to errors that look like model hallucinations but are actually compatibility issues. [1 ]
Missing Features: OpenCode is built to be provider-agnostic; it normalizes differences (e.g., tool call IDs, varying context limits) across models. Claude Code assumes it is talking to Claude; if the proxied model doesn’t support specific “bells and whistles” (like specific caching or reasoning formats), the session can become unstable. [1 , 2 , 3 , 4 , 5 ]

Performance & Speed (Mixed)

Agent Efficiency: While a proxy adds negligible network latency (microseconds), the workflow efficiency drops. Users report that non-Claude models inside the Claude harness handle the context window differently, leading to a different “cadence” and often requiring more manual correction. [1 , 2 , 3 ]
Context Management: Claude Code aggressively optimizes context (e.g., compacting history) in a way that non-Claude models may not handle well, potentially causing “context rot” or loss of critical information during long sessions. [1 , 2 ]