Wikit/RoutingCompendium-cost

RoutingCompendium — Cost

Inference price of every candidate LLM appearing in Wikit/RoutingCompendium-perf. The two datasets are meant to be loaded together: -perf gives what each candidate scores on a query, -cost gives what calling it costs.

Splits

One split per benchmark, with the same names as RoutingCompendium-perf (RouterBench, Sprout, EmbedLLM, FusionBench, R2Bench). Each split lists the candidates of that benchmark's pool — a few dozen rows at most.

Schema

One row per candidate model.

Field	Type	Description
model	string	Model name, matching an entry of models_name in the corresponding -perf split.
cost	string	Price, stored as a string and parsed with ast.literal_eval.

The two shapes of cost

cost is a string so that one column can hold both a scalar and a dict. After ast.literal_eval you get either:

1. A dict of per-token prices — {"input": float, "output": float}, in USD per 1M tokens, as published by the provider. Used by RouterBench, FusionBench and R2Bench.

{'input': 0.09, 'output': 0.55} 

**2. A number** — the model's parameter count, in billions. Used by `Sprout` and `EmbedLLM`.

```python
7.0    # a 7B model

Parameter count -> USD per 1M tokens

Open-weight models are priced by size, using the Together AI pricing as accessed on 2025-06-05:

Parameters (B)	USD / 1M tokens
≤ 4	0.10
≤ 8	0.20
≤ 21	0.30
≤ 41	0.80
≤ 80	0.90
≤ 110	1.80
> 110	1.80 + 0.03 × (params − 110)

Beyond 110B the last interval's slope ($0.03 per additional billion parameters) is extrapolated linearly.