How we prove it
Four promises, each one independently verifiable. If any check fails, your billing period is refunded — no ticket, no arguments.
1. One model name → one official upstream
claude-* goes to Anthropic. gpt-* goes to OpenAI. gemini-* goes to Google. No distilled clones, no silent downgrades under load.
2. No prompt injection, no response rewriting
Your request body is forwarded byte-for-byte. The only edits are auth key swap and (optional, opt-in) schema translation between OpenAI and Anthropic protocols.
3. Daily benchmark parity, published raw
Every 24 hours we run MMLU-Pro / GSM8K / HumanEval samples via VoltAI and via direct upstream. We publish the delta. Target ≤ 1.5%.
4. Bring your own key (zero-trust)
Enterprise plans let you configure your own upstream provider keys. VoltAI then only handles auth, metering, rate-limiting — we never see the model output.
Today's parity snapshot
Auto-generated at 00:05 UTC. Numbers below are current live readings. Raw eval logs at github.com/voltai/integrity-eval.
| Model | Upstream | MMLU-Pro | GSM8K | HumanEval | Δ vs direct | Status |
|---|---|---|---|---|---|---|
| DeepSeek API | DeepSeek | — | 90.0% / 90.0% | — | +0.00pp | pass |
| Claude family | Anthropic | — | — | — | — | pending |
| ChatGPT / GPT family | OpenAI | — | — | — | — | pending |
| Gemini family | — | — | — | — | pending | |
| Open-source deployments | VoltAI-owned hardware | — | — | — | — | pending |
Don't trust us — verify us.
Grab $5 free credit and run your own eval. Our public eval repo reproduces every number you see above. If our numbers don't match yours, we refund.