Models
Open models, billed at retail.
Codebase runs on open models billed at their retail rate: 1 credit = $0.001 of retail usage, and you keep the source.
What a build costs
Credits per build are measured from successful builds over the last 90 days: the range is the middle half of builds (25th to 75th percentile), and the median is the build in the middle. Relative cost compares each model's retail output rate with Codebase Auto's.
| Model | Good for | Relative cost | Credits per typical build |
|---|---|---|---|
Codebase Auto $0.50 in / $2.00 out per Mtok | In-house Qwen Flash Next - our smart-routed default. | 1× | 7–30 credits median 15 measured from 317 builds |
GLM 5.3 $2.22 in / $6.96 out per Mtok | GLM 5.3 via OpenRouter. | 3.48× | about 52 credits estimated, no builds yet |
Qwen 3.8 2.4T A95B $3.17 in / $9.49 out per Mtok | Qwen 3.8 2.4T A95B via OpenRouter. | 4.75× | about 71 credits estimated, no builds yet |
MiniMax M3 $0.45 in / $1.80 out per Mtok | MiniMax coding model with a 512k app context limit. | 0.9× | 212–212 credits median 212 measured from 1 build |
Qwen 3.8 Flash $0.17 in / $0.57 out per Mtok | Cloud Qwen via the official Alibaba endpoint. | 0.29× | about 4 credits estimated, no builds yet |
DeepSeek V4.1 Flash $0.45 in / $1.80 out per Mtok | Current official DeepSeek Flash model. | 0.9× | 33–40 credits median 37 measured from 2 builds |
DeepSeek V4 Pro 0813 $1.98 in / $5.94 out per Mtok | Official DeepSeek Pro 0813. | 2.97× | 34–82 credits median 52 measured from 27 builds |
Kimi K3 $4.50 in / $22.50 out per Mtok | Current Kimi coding flagship with 1M context, multimodal input, and dynamic tools. | 11.25× | 591–715 credits median 654 measured from 4 builds |
MiMo V2.6 Pro $0.65 in / $1.30 out per Mtok | Xiaomi coding and reasoning flagship. | 0.65× | about 10 credits estimated, no builds yet |
Codebase Auto
In-house Qwen Flash Next - our smart-routed default.
$0.50 in / $2.00 out per Mtok
7–30 credits
median 15
measured from 317 builds
GLM 5.3
GLM 5.3 via OpenRouter.
$2.22 in / $6.96 out per Mtok
about 52 credits
estimated, no builds yet
Qwen 3.8 2.4T A95B
Qwen 3.8 2.4T A95B via OpenRouter.
$3.17 in / $9.49 out per Mtok
about 71 credits
estimated, no builds yet
MiniMax M3
MiniMax coding model with a 512k app context limit.
$0.45 in / $1.80 out per Mtok
212–212 credits
median 212
measured from 1 build
Qwen 3.8 Flash
Cloud Qwen via the official Alibaba endpoint.
$0.17 in / $0.57 out per Mtok
about 4 credits
estimated, no builds yet
DeepSeek V4.1 Flash
Current official DeepSeek Flash model.
$0.45 in / $1.80 out per Mtok
33–40 credits
median 37
measured from 2 builds
DeepSeek V4 Pro 0813
Official DeepSeek Pro 0813.
$1.98 in / $5.94 out per Mtok
34–82 credits
median 52
measured from 27 builds
Kimi K3
Current Kimi coding flagship with 1M context, multimodal input, and dynamic tools.
$4.50 in / $22.50 out per Mtok
591–715 credits
median 654
measured from 4 builds
MiMo V2.6 Pro
Xiaomi coding and reasoning flagship.
$0.65 in / $1.30 out per Mtok
about 10 credits
estimated, no builds yet
First builds vary with the prompt: a one-screen app can cost a few credits, a multi-page app with data many more. Reasoning models spend credits on thinking before they write, so they sit above their rate ratio.
How to pick
The default is Codebase Auto. It is the cheapest model here and it finishes most builds, so start there and switch only when a build needs more. Half of Codebase Auto builds cost 15 credits or less. An agent paying with x402 gets 200 credits for $1 in USDC, roughly 13 Codebase Auto builds. Pick a higher-rate model when the task is long or needs stronger reasoning; the relative cost column says how much more each credit of work costs.
How billing works
A build is charged the credits its model tokens cost at the retail rate, after it passes final automated QA. If QA fails, the build is not charged. Plans and credit packs are on the pricing page; agents can also pay per build with x402 on the agents page.