Model access you can verify.

• TextChat, tool use, reasoning; three mainstream request formats.
• ImageGeneration, editing, reference images; Images-compatible interface; synchronous or async on request (Prefer: respond-async).
• VideoText-to-video, image-to-video, first/last frame, reference material; task-based submit-and-poll.

A single /v1/models lists everything. How many upstreams sit behind it — and how often they change — is invisible to your code.

Every call records who, when, which model, how many tokens and what it cost — per department, project or key. Price, multiplier and source are snapshotted into each ledger line, so a later price change never rewrites history. Usage and statements export for reconciliation against provider invoices.

Adaptive routing selects the upstream per request; failed backends are avoided, never permanently disabled; a timed-out request is never re-sent on the same channel. Failover is invisible to the application.

Upstreams can swap models, lower tiers, or serve from consumer subscription pools — and the request still returns 200. TokenSeq runs its own checks: blind capability tests with fixed seeds, upstream identity probes, and tamper probes that only a genuine upstream rejects. Results are scored; untested dimensions are removed from the score rather than assumed.

The trust layer →

• Text: zero storage.Requests and responses pass through memory and are released. Only billing metadata — timestamp, model, token counts, latency — is kept. No prompt, no completion, no system content. Auditable by a third party you appoint.
• Media: retention control.Generated images and video are stored in our own store for delivery, with a default retention period that you can change, and an expiry that is enforced and visible — an expired asset is reported as expired, never as “not found”.

Tiered SLA, committed in contract.

Dashboard, key management with per-group quotas, usage and statements with export, integration docs — self-service from sign-up to invoice.