Tiers: pick your altitude
In one breath: not every workload needs the strongest guarantee, and paying for one you don't need is its own kind of waste. The guarantee comes in three strengths, and you set it per call.
The mountain
We name the tiers after altitude because that's how they feel in practice: the higher you climb, the stronger the guarantee and the more it costs in latency.
Tier 1 · Foothill · Standard
Fast access to a wide choice of large models, with secrets scrubbed on your device before a request ever leaves it. Zero retention, lowest cost, and yes, still a receipt on every call.
For: general workloads where the input isn't regulated data, but you still want hygiene and evidence.
Tier 2 · Ridge · Private
The regulated default. The compiler picks the protection that fits your anchor and workload, from masked serving (the serving node never sees your raw inputs) to attested enclaves. Benchmarked below 0.05% token recovery, at interactive speed.
For: production workloads on regulated data that need answers now.
Tier 3 · Summit · Sovereign
The strongest guarantee we offer. Your data is split into pieces and sent to operators who don't know each other. No single one of them holds enough to reconstruct anything, and the result comes back with a STARK proof of correct execution.
Two details your security team will like:
- The splitting itself can happen on your own device. The "trusted dealer" role in MPC doesn't have to be a server you rent from us. Shares can be created inside your walls before anything leaves.
- The privacy is information-theoretic under non-collusion. That's the same standard assumption behind MPC everywhere: the guarantee holds as long as the operators don't all conspire.
For: batch workloads where assurance beats latency. You trade speed for proof.
Side by side
| Foothill · Standard | Ridge · Private | Summit · Sovereign | |
|---|---|---|---|
| Speed | Interactive | Interactive | Batch |
| Who sees raw inputs | Scrubbed on-device first | No serving node, ever | No one; data is split |
| Token recovery | n/a (scrubbed) | < 0.05%, benchmarked | 0% effective |
| Proof | Receipt | Receipt | Receipt + STARK |
| Typical use | General workloads | Regulated production | Highest-assurance batch |
Changing your mind is one field
The tier is a parameter on the request, not a contract negotiation:
{ "task": "vision.classify", "tier": "sovereign" }
Run your nightly compliance batch at Summit and your customer-facing chat at Ridge, from the same API key.
Standard and Private tiers serve production traffic. The Sovereign tier's MPC serving path is active R&D: it runs, we benchmark it continuously, and we deploy it per engagement with the caveats agreed in writing. Recovery-rate figures are our own benchmarks, not third-party certifications. See the Status board and Disclosures.
Next: decide where it runs →