Use them to give your agents judgement and taste
A face tells a model who so it can say why. They work with existing LLMs and our native decision models. Use a face as the voice of your agent, or give it access to the Faces cli and it can spool up networks of faces for bespoke special intelligence on demand.
1000x faster
3000x cheaper
Than fine tuning
Faces are compiled, not trained. They use a proprietary form of mechanical interpretability that we call a Fast Fourier Transform for Natural Language, or FNT. The Faces compiler uses FNT to break down sample texts (and, soon, images too) into the component parts of a mind and then uses them to steer LLMs and decision models.
Compiling with the FNT makes faces composable and comparable. The Faces API provides endpoints for doing boolean algebra with faces and computing differences between their centroids in the various bands of cognitive latent space.
The future of intelligence is special, not general. A mind that belongs to no one, weighing every question from no particular place, knows nothing in particular. Intelligence requires more than context, it needs a point of view. Use faces.
Research at Headwaters Design Lab
Pricing
$17/mo
Make unlimited faces and Application Specific Intelligence Circuits using System One and Two models.
Each System One model can give three kinds of type-cast response: bool, choice and score. A System One model can be bare or wear a face; the cost is the same either way. Faces offers two tiers.
| Tier | Input | Output |
|---|---|---|
| Large | $0.042 | Free1 |
| Small | $0.012 | Free1 |
A System Two model is a large language model with the face you choose applied to it, so it takes on that face’s perspective. Inference costs the underlying provider’s price plus a 5% markup.
Or connect an OpenAI subscription over OAuth: Faces then runs on OpenAI’s Codex models through your own plan, and you don’t pay per token.
| Model | Provider in / out | Faces in / out |
|---|---|---|
| GPT-5OpenAI | $1.25 / $10.00 | $1.31 / $10.50 |
| GPT-5 mini | $0.25 / $2.00 | $0.26 / $2.10 |
| o4-mini | $1.10 / $4.40 | $1.16 / $4.62 |
| Claude Opus 5.5Anthropic | $15.00 / $75.00 | $15.75 / $78.75 |
| Claude Sonnet 5 | $3.00 / $15.00 | $3.15 / $15.75 |
| Claude Haiku 4.5 | $1.00 / $5.00 | $1.05 / $5.25 |
| Gemini 2.5 ProGoogle | $1.25 / $10.00 | $1.31 / $10.50 |
| Gemini 2.5 Flash | $0.30 / $2.50 | $0.32 / $2.63 |
| DeepSeek-V3DeepSeek | $0.27 / $1.10 | $0.28 / $1.16 |
| DeepSeek-R1 | $0.55 / $2.19 | $0.58 / $2.30 |
| Qwen3 235BQwen | $0.20 / $0.60 | $0.21 / $0.63 |
| Qwen3 32B | $0.10 / $0.30 | $0.11 / $0.32 |
| MiniMax-M1MiniMax | $0.40 / $2.20 | $0.42 / $2.31 |
| GLM-4.5Z.ai | $0.60 / $2.20 | $0.63 / $2.31 |
| GLM-4.5 Air | $0.20 / $1.10 | $0.21 / $1.16 |
Provider prices shown are illustrative.
Select task to try Faces
Configure circuit
Faces lets you build Application Specific Intelligence Circuits, cast them to an API endpoint, and call them whenever you need.
Each circuit is made of one or more System One and System Two models: fast, cheap decision makers, and text-generating language models, respectively. Each can be bare, or wear the face of a particular person or type of person. Train one on your own emails and messages to be your communication surrogate, and another on an ancient philosopher for wisdom on demand.
You’ll choose
Chosen
Your circuit, as one call
POST https://api.faces.sh/api/v1/circuits/
Send it any prompt, from anywhere. It runs every module and face you chose, in order.
| Module | Model | Input | Output |
|---|---|---|---|
| S1·A | System One | $0.042 | Free |
| S1·B | System One | $0.042 | Free |
| S1·C | System One | $0.042 | Free |
| S2·A | $0.000 | $0.000 | |
| S2·B | $0.000 | $0.000 | |
| Circuit | every module, once | $0.126 | $0.000 |
System Two at the provider’s price plus 5%; System One output too cheap to meter. Illustrative rates.
0:00 / 0:00