How caft works

Routing

caft forwards each request to an upstream provider that serves the model you asked for. You use one key and one balance for every model, and any OpenAI-compatible client or router works.

Upstream providers are not the model makers. caft checks quality with its own test requests, never by reading your traffic. Capacity depends on those providers, so a model can be briefly unavailable (HTTP 503).

Prices

Prices aim for at least 75% under the maker's list price and half the cheapest OpenRouter offer. See Pricing and credits.

Failures and retries

A request that fails before any output is not charged. A stream that fails partway is charged only for the tokens delivered. On 429, 502, 503 or 504, retry with backoff, or let your router fall back to another provider. See Troubleshooting.

There is no uptime commitment. Live success rates, time to first token and throughput for each model are on the status page.

Data

• caft keeps no prompts or completions. Content is held in memory only while the request runs.
• caft logs metadata for billing: time, model, token counts, duration and cost.
• Upstream providers have their own policies and may keep what they receive. Don't send personal data.

Details are in the Privacy Policy.