Pay for models, not a mystery abstraction.
One inference surface, no hidden economics. Bring your own keys for zero Sofia charge, or use hosted models with published, metered pricing.
One model ID, several ways to run.
A Sofia model can prefer an org key, fall back to PAYG, or route to hosted infrastructure. Routing, health, and pricing stay visible.
Your key stays useful.
If an organization credential exists, Sofia can use it first. A depleted Sofia balance never blocks a working BYOK route.
One account when you want it.
No provider setup required. Sofia meters the exact request, settles the wallet, and records provider cost separately from customer price.
Pricing beyond input and output tokens.
Providers now price cache reads, reasoning, search, media, batch jobs, and time windows differently. Sofia’s price books model those rules explicitly.
Cached vs uncached
Separate meters preserve cache economics instead of flattening everything into one input token number.
Peak and off-peak
Versioned price rules can change by UTC window without hardcoding provider policy into request code.
Request-based charges
Web search and other server tools can be priced per request or any arbitrary unit size.
Audio and images
Preflight estimates and final usage keep multimodal requests inside the same reserve/settle lifecycle.
Immutable line items
Every request stores the price-book version, meter quantities, unit rates, provider cost, and customer charge.
Provider statements
Automated reconciliation matches upstream statements to Sofia requests and highlights unexplained variance.
Point existing clients at Sofia.
The public inference surface follows familiar `/v1` conventions, while Sofia keeps richer routing and metering behind the endpoint.
curl https://api.sofia.example/v1/responses \
-H "Authorization: Bearer sk-sofia-…" \
-H "Content-Type: application/json" \
-d '{
"model": "sofia/deepseek-v4.1",
"input": "Review this patch"
}'Bring a key or bring nothing.
BYOK remains first-class. Sofia PAYG is there when you want one balance, managed routing, and a team-wide control plane.