# CN-Open Basket — Methodology

> Index code: `CN-OPEN`  
> Type: Themed fixed compute price index  
> Frequency: Weekly  
> Status: Published v0.2 (frozen groups in `scripts/build_cn_indices.py`)

---

## 1. Purpose

Measure the hosting-cost centre for **open-weight models** on open inference hosts — the open inference deflation curve — and contrast it with the Frontier closed basket.

---

## 2. Universe

### 2.1 Models

Primary key: `compare_group_id`. Include only if:

- `model_source = open` (or equivalent open label)
- Belongs to a maintained mainstream family (e.g. Llama, Qwen, DeepSeek, GLM, Mistral, Gemma)
- At least **two** distinct `compute_provider`s have eligible chat×token quotes that week (comparability; less single-host noise)

### 2.2 Providers (default whitelist)

Prefer: DeepInfra, Together AI, Fireworks AI, Groq, SiliconFlow, OpenRouter (only when routing an open upstream).

May exclude: pure resellers; quotes > 5× the group median (obvious outliers).

### 2.3 Initial representative groups (freeze before launch)

One or two active `compare_group`s per family (e.g. DeepSeek V4 Flash/Pro, Qwen3 flagship MoE, current Llama instruct). Target **8–12** groups; avoid over-concentration in one lab.

---

## 3. Constituent price

For open group $j$:

$$
p_{j,t} = \mathrm{median}\{\text{blended quotes in group } j\}
$$

blended = $0.7\cdot\text{input}+0.3\cdot\text{output}$.

---

## 4. Basket price (primary unit: USD)

Same as Frontier: equal-weight geometric mean over active groups.

$$
P_t = \exp\left(\frac{1}{n_t}\sum_j \ln p_{j,t}\right)
$$

**Headline unit: USD / 1M tokens**. Do not publish if $n_t < 5$.

### Optional: index points

$$
I_t = 100 \times \frac{P_t}{P_{t_0}}, \quad t_0 = 2026\text{-}06\text{-}20
$$

---

## 5. Rebalance

- **Light monthly review**: drop groups with no hosts; new families may enter after two consecutive weeks with ≥2 providers.
- **Formal quarterly rebalance**: freeze constituent version (e.g. `CN-OPEN-2026Q3`).

---

## 6. Limitations

- Open model IDs churn faster than Frontier; higher constituent drift.
- Quantisation / throughput differences across hosts are not in the price (list-price index).
- China-region and global hosts are pooled; geo splits belong in cross-signals.

---

## 7. Relation to CN-Token Spot / Frontier

| Index | Question answered |
|-------|-------------------|
| Token Spot | How cheap is the curated market overall? |
| Frontier | How expensive is top closed intelligence? |
| Open | How are mainstream open hosting costs moving? |