Data & security
We train our optimization engine, not a model.
Atlas learns which route and model wins on cost and quality — it never learns from the contents of your prompts. Your prompts and code are never absorbed into a language model, never sold, never shared. Your IP stays yours. This page is the plain-English version; the binding terms live in our Privacy Policy and Terms of Service.
The distinction that matters
An engine that picks routes — not a model that reads your code.
Most people's first question is whether we train on their data. Here is the honest answer: we train the part of Atlas that decides which model and route serves a call the cheapest way that holds quality. That decision engine learns from outcomes — cost, latency, and whether quality held. It does not learn from what your prompts say.
What the engine learns from
- Which route or model won on cost and quality for an operation
- Cost and price per call, so savings stay verifiable line-by-line
- Latency and reliability of each path on your traffic
- Quality signals — evaluator scores, pass/fail, and your thumbs-up / thumbs-down
What it never does
- Train a language model — shared or otherwise — on your prompts, code, or outputs
- Sell your prompts, code, or outputs to anyone, ever
- Share your content with other customers or third parties to build a product
- Use your content to fine-tune a per-tenant model unless you opt in to the Reliability Loop
Your controls
On by default. Off in one click. Isolated on Enterprise.
By default we use your traffic to improve routing, cost, and quality — that's how your savings climb as Atlas ramps. You're never locked into that.
Default: usage-based optimization
We use your routing outcomes — cost, latency, quality — to tune how Atlas serves your workload. Your content is processed to deliver and improve the Services for you, never to train a language model.
Free per-key off-switch
Turn off usage-based optimization per API key, at no cost. Atlas still optimizes each call in real time and you still get the price and quality guarantees — we just stop using that key’s traffic to tune routing over time.
Enterprise: full data isolation
Enterprise plans get full data isolation — your traffic is walled off from any cross-customer optimization, with terms tailored to your compliance needs.
Opt-in only
Per-tenant tuning is a deliberate choice — never a default.
If you want a model variant tuned to your own workload, the Reliability Loop lets you build golden datasets and request tuning on data you explicitly mark and submit. It's opt-in, it's governed by its own terms, and any resulting variant is yours alone — never shared with other customers.
Nothing in default usage-based optimization fine-tunes a model on your content. Per-tenant tuning only happens when you opt in through the Reliability Loop. See Terms §5 for the governing language.
Retention & subprocessors
How long we keep payloads, and who helps us run the Services.
We retain API request and response payloads for the window set by your plan — typically 30 or 90 days — then delete them automatically. You can request earlier deletion any time at privacy@newmen.ai.
Named subprocessors
- Hosting and infrastructure: Vercel, Neon (database)
- Payment processing: Stripe
- Transactional email: Resend
- Model inference: OpenRouter (API request proxying)
- Analytics: privacy-preserving analytics that does not fingerprint users across sites
We notify customers at least thirty (30) days before a new subprocessor begins processing data. The current, complete list is available on request at privacy@newmen.ai. Full detail lives in the Privacy Policy.