Deep Agents Open Source Open Models have crossed a threshold
Swap to an open model like GLM‑5 or MiniMax M2.7 in Deep Agents by passing the model string to create_deep_agent, benefiting from lower cost and comparable performance.
Install the appropriate model provider SDK, then create a Deep Agent with model='baseten:zai-org/GLM-5' or 'openrouter:minimax/minimax-m2.7', and run your agent to benefit from lower cost and comparable performance.
Summary
Open‑weight models such as GLM‑5 (z.ai) and MiniMax M2.7 (OpenRouter) now match closed‑frontier models on core agent tasks like file operations, tool use, and instruction following, while offering lower cost and latency. The article presents evaluation results across seven categories (file ops, tool use, retrieval, conversation, memory, summarization, unit tests) using metrics of correctness, solve rate, step ratio, and tool call ratio. GLM‑5 achieves 0.64 correctness (94/138), solve rate 1.17, step ratio 1.02, and tool call ratio 1.06; MiniMax M2.7 scores 0.57 correctness, 0.27 solve rate, 1.02 step ratio, and 1.04 tool call ratio. Pricing tables show GLM‑5 at $0.95/M input and $3.15/M output tokens, MiniMax M2.7 at $0.30/M input and $1.20/M output, compared to Claude Opus 4.6 at $5.00/M input and $25.00/M output. Latency comparisons reveal GLM‑5 at 0.65s and 70 tokens/s versus Claude Opus 2.56s and 34 tokens/s. Deep Agents harness automatically injects model identity, adapts context limits, and handles unsupported modalities. The CLI supports ConfigurableModelMiddleware for mid‑session model switching via a /model command.
Key takeaways include the viability of open models for production agents, the cost savings, and the harness’s ability to adapt to different model characteristics.
Future work will explore multi‑model subagent configurations and further tuning patterns for specific open model families.
Key changes
- GLM‑5 and MiniMax M2.7 achieve similar correctness to closed models on core agent tasks
- GLM‑5 costs $0.95/M input, $3.15/M output; MiniMax M2.7 costs $0.30/M input, $1.20/M output
- GLM‑5 latency 0.65s, 70 tokens/s vs Claude Opus 2.56s, 34 tokens/s
- Evaluation metrics: correctness, solve rate, step ratio, tool call ratio
- GLM‑5 correctness 0.64, solve rate 1.17, step ratio 1.02, tool call ratio 1.06
- MiniMax M2.7 correctness 0.57, solve rate 0.27, step ratio 1.02, tool call ratio 1.04
- Deep Agents harness auto‑injects model identity, context management, and adapts to model limits
- ConfigurableModelMiddleware allows mid‑session model switching via /model command