FAQ

POST /v1/chat/completions (OpenAI-compatible) and POST /v1/messages (Anthropic-compatible). Authenticate with Authorization: Bearer sk-bhaskara-… (Anthropic-style x-api-key works on /v1/messages). Models you can name: glm-5.3, qwen-3.8, theta. The flash variants of the frontier families exist upstream — our router selects them per-session; you address them through the two frontier endpoint names.

A flat per-request endpoint — you stop counting tokens, we absorb the variance. It routes your turn to the cheapest backend that can handle it: routine turns to flash-tier capacity, turns our classifier flags as heavy reasoning to a stronger backend. If your agent is chatty and context-heavy, theta is usually the cheaper bill.

Never silently. A request naming glm-5.3 is answered by the GLM-5.3 family. Which variant — full or flash — is the router's session-level decision, disclosed in the response and headers (see the sticky lock below). You always know which endpoint you called and which tier answered.

Provider prompt caches are keyed to a stable request prefix and model. If we flipped your session between a full model and its flash variant turn to turn, every flip would wipe the cache and re-bill your prefix. So the first request of a session gets routed by task difficulty — and the tier chosen for that session then sticks (2-hour idle TTL). One priced exception: a genuinely hard turn (debugging, planning) in a flash-locked session can trigger a one-time upgrade to the full model — but only when the cache-wipe penalty (re-billing your prefix at full rates) stays under a strict budget and your weekly full-model share has headroom. Fresh sessions get fresh decisions.

Put the stable stuff first: system prompt, tool definitions, long-lived context. Append only at the end. Any mid-context edits invalidate everything after them. Watch x-ratelimit-remaining-* and your dashboard cache column; we publish the exact prefix rules the router assumes.

Every successful response: x-quota-plan, x-quota-monthly-reset; theta responses add x-ratelimit-limit-requests / -remaining-requests; frontier responses add x-ratelimit-limit-tokens / -remaining-tokens and x-quota-output-tokens-remaining. On 429 you additionally get Retry-After in seconds. No pricing data is ever in headers — costs live in the dashboard.

Frontier endpoints count prompt and completion tokens separately against monthly caps (reset 00:00 UTC on the 1st). Theta counts requests in a rolling five-hour window plus a monthly ceiling. Exceed anything and you get a clean 429 with the reset time — never a silent degrade.

Rejected with 429 and a dashboard deep-link; the sticky lock for that session stays locked — after reset, your session resumes on the same model with the cache still warm.

Cohorts of 100. New users land on the waitlist while a cohort review is running so existing users keep stable capacity; the gate is automatic and admin-dialled.

Only while your dashboard "Train on my traffic" switch is on — flip it and the very next request stops being stored. Full mechanics in the training disclosure. The honest limit of opt-out (gradients don't unlearn) is stated there too.

India-hosted primary store, encrypted at rest; upstream inference providers see the live request only and are contractually zero-retention for pass-through traffic. Details in the privacy policy.

Because it is the only honest comparison: the same uncached prompt at full list price of the frontier family, no cache discounts. It is an estimate, labelled as one — calculator is public, and we disclose the three engineering choices (cache engineering, context management, routing) that make our prices possible.

7-day full refund on a first paid cycle; pro-rata on downtime > 24h and billing errors — see the refund policy.