OptiLLM's approaches are token multipliers — bon ×5, moa ×8, mcts can be ×20. That's exactly why base model cost-per-token matters so much here: running moa on a frontier-priced model gets expensive fast, but running it on a cost-efficient model is where the "small model beats big model" magic becomes practical.

I built AnyLLM, an OpenAI-compatible API gateway with 24+ Chinese and international models behind a single key: DeepSeek V4, Qwen3.7-Plus, GLM-5.2, Kimi K2.5, MiniMax M3 — alongside GPT-5 and Gemini. International cards and crypto accepted, pass-through pricing with zero markup, built-in routing and failover.

Why this fits OptiLLM
Works today, zero integration. AnyLLM is just an OpenAI-compatible endpoint — set base_url and ANYLLM_API_KEY, then run bon-deepseek-v4, moa-qwen3.7-plus, mcts-kimi-k2.5 out of the box.
The cost math is compelling. Your own benchmarks show techniques like MOA lift GPT-4o-mini to GPT-4-level. Applying the same techniques to DeepSeek or MiniMax pricing could make optimized small-model pipelines nearly free compared to frontier models.
One key for the whole benchmark matrix. For future eval runs, AnyLLM gives you one signup covering all major Chinese frontier models — no separate accounts, phone numbers, or Alipay.
Synergy with the proxy plugin. AnyLLM's built-in routing/failover plus OptiLLM's proxy plugin makes a solid production stack: optimization layer on top, multi-provider resilience below.
Proposal
A short docs section or provider example: "Using an OpenAI-compatible gateway (AnyLLM)".
If useful for your eval work, I'd be glad to provide a sandbox key so you can benchmark Chinese models with your approaches — results like "CePO on DeepSeek V4" would be genuinely novel data points.
(Context: I'm the developer of AnyLLM.)

Best,
Leo Bennett
[leo.indiedev@gmail.com]