The allowlist, the baseline, and routing
You pick from a fixed set of models — and you may route different tasks to different ones. Model selection and routing is part of what you're optimizing for cost.
The allowlist
| Model | Tier | Notes |
|---|---|---|
z-ai/glm-5.2 | baseline | the bar to beat |
deepseek/deepseek-v4-pro | frontier | deep reasoning, pricier |
deepseek/deepseek-v4-flash | fast | cheap, quick turns |
minimax/minimax-m3 | frontier | long-context coder |
xiaomi/mimo-v2.5-pro | frontier | high-effort tasks |
xiaomi/mimo-v2.5 | mid | balanced cost/quality |
stepfun/step-3.7-flash | fast | budget throughput |
openai/gpt-oss-120b | open | open-weight generalist |
Free model variants are banned — they distort billed cost. Kimi K3 is supported by the pinned pi build but temporarily deferred while its provider returns overload errors; it is not currently selectable or scorable.
Routing is a lever
You aren't locked to one model. A good harness sends cheap, mechanical tasks to a fast model and reserves an expensive frontier model for the hard ones — that mix is often what separates the cheapest qualifying entry from the rest.
┌─────────────────────────┐
incoming task ───────▶ │ your routing logic │
└───────────┬─────────────┘
simple / mechanical? │ hard / long-horizon?
▼ │ ▼
┌───────────────────┐ │ ┌────────────────────┐
│ step-3.7-flash │ │ │ deepseek-v4-pro │
│ deepseek-v4-flash │ │ │ minimax-m3 │
│ (cheap, fast) │ │ │ (costly, capable) │
└───────────────────┘ │ └────────────────────┘
└──────────────┬───┴────────────────┘
▼
lowest billed $ for the pass count you hit
The baseline
The baseline harness is vanilla pi on z-ai/glm-5.2. Its pass count sets the eligibility bar and its billed cost is the diamond on the frontier chart. Beat its pass count to qualify; beat its cost to climb.