Models

One allotment, many models, a published conversion between them. Every performance number here carries the date it was measured, and the ones that have not been measured say so.

What we found by testing these models, and the probe behind each finding, on /docs/why

Catalog

5 models being served

Send the slug as the model field. It is the same string on every surface: activity rows, response headers, and your config file.

glm-5.3-flash

GLM-5.3 Flash · Z.ai

×0.05

serving

Multiplier: ×0.05
Context served: 1.00M tokens
Max output: 32.8K tokens

in ×1.0 · cached ×0.2 · out ×3.333333

deepseek-v4-flash

DeepSeek V4 Flash · DeepSeek

×0.31

serving

Multiplier: ×0.31
Context served: 1.00M tokens
Max output: 32.8K tokens

in ×1.0 · cached ×0.031818 · out ×3

deepseek-v4-pro

DeepSeek V4 Pro · DeepSeek

×0.94

serving

Multiplier: ×0.94
Context served: 1.00M tokens
Max output: 32.8K tokens

in ×1.0 · cached ×0.033333 · out ×3

glm-5.3

GLM-5.3 · Z.ai

×1.0

serving

Multiplier: ×1.0 (reference)
Context served: 1.00M tokens
Max output: 32.8K tokens

in ×1.0 · cached ×0.185714 · out ×3.142857

kimi-k3

Kimi K3 · Moonshot

×2.1

serving

Multiplier: ×2.1
Context served: 1.00M tokens
Max output: 32.8K tokens

in ×1.0 · cached ×0.1 · out ×5

Weighting

What a multiplier is

Models cost materially different amounts to serve, so one allotment is converted into each of them at a published rate. A cheaper model makes your allotment go further; that is a real benefit, not a footnote.

weighted tokens

WT = (in × W_IN + cached × W_CACHED + out × W_OUT) × M_model

Model	M	W_CACHED	W_OUT	1M WT buys, in output tokens
glm-5.3-flash	×0.05	×0.2	×3.333333	5,600,001
deepseek-v4-flash	×0.31	×0.031818	×3	1,060,605
deepseek-v4-pro	×0.94	×0.033333	×3	353,535
glm-5.3 · reference	×1.0	×0.185714	×3.142857	318,182
kimi-k3	×2.1	×0.1	×5	93,333

The last column is pure output, which no real workload is. It is there as a common yardstick between rows, not as a quota estimate. W_IN is ×1.0 on every model by definition; the multiplier carries the difference between them.

A worked request on glm-5.3. Your own requests show this same arithmetic, with your real numbers, in the activity detail.

one request

12,400 in, of which 3,180 were cached

9,220 uncached in  × 1.0       =   9,220.00
3,180 cached       × 0.185714  =     590.57
8,920 out          × 3.142857  =  28,034.28
────────────────────────────────────────────
subtotal                       =  37,844.85
× 1.0 (glm-5.3)                =     37,845 WT

Cached tokens are counted as part of the input, not on top of it, so the input line bills the uncached remainder: the cache discount is real, not a discount on a number we inflated first. Rounding to whole weighted tokens happens once, on the final line; that is why the lines above carry decimals.