Developer uses GLM-5.3-Flash and GPT-6.1 Sol for coding
A developer reports using a self-hosted GLM-5.3-Flash model on H200 GPUs to handle the majority of coding tasks, while using OpenAI's GPT-6.1 Sol for high-level planning and review. The setup aims to reduce costs and API usage limits associated with premium models like Claude Opus 5.5 or GPT-6 Astra. The developer notes that while frontier models are superior for complex tasks, GLM-5.3-Flash is…
Key points
- Developer uses GLM-5.3-Flash on H200s for coding, with GPT-6.1 Sol for planning and review.
- OpenCode harness enables routing subagents to local models, unlike Claude Code or Codex.
- Test task sent 60% of tokens to GLM; larger tasks estimated at 70-75% local usage.
The workflow relies on OpenCode, a coding harness that allows different agents to use different model providers. In this configuration, Sol acts as the orchestrator, writing briefs and verifying results, while GLM agents handle code reading, editing, and testing. The developer found that Claude Code and Codex do not support routing subagents to local endpoints, making OpenCode or Pi necessary for this hybrid approach. Permissions are strictly managed to prevent the local model from executing unauthorized shell commands.
In a test task, approximately 60% of tokens were processed by GLM, with the developer estimating that larger tasks would shift 70-75% of token usage to the local model. This setup allows the team to share the GPU server, with requests queued during busy periods. The developer emphasizes that this pattern is not dependent on GLM specifically and could work with other open-source models like Qwen3.8-27B, provided the harness supports multi-provider routing.
Model pages: GPT-6.1 Sol → · Claude Opus 5.5 →
GLM-5.3-Flash on our H200s does most of my coding now while GPT-6.1 Sol orchestrates (OpenCode setup)
aayushgarg.dev · 1 November 2026
Loading the full article…
This text was published by aayushgarg.dev. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Microsoft tests Context Vault, a new AI memory for Windows · 1 src
- Karpathy ships HackerTube in ~2 hours with Claude Opus 5.5 · 1 src
- Pharmacist moves Claude Code daily tasks to Windows Task Scheduler · 1 src
- DeepSeek Harness team releases open-source Super-agent prototype · 4 src
- User shares Claude Code workflow on Mac mini with manual approval limits · 1 src
Comments
via GitHub Discussions