Smart routing
Each call goes to the cheapest model that meets your quality bar — simple tasks on DeepSeek/Kimi, hard ones on GPT/Claude only when required.
We audit your real API traffic and route each task to the cheapest model that still meets your quality bar — DeepSeek, Kimi, or Qwen instead of GPT or Claude. Same prompts. Same output. 70–90% less — with data-residency controls you choose.
BYOK by default · prompts never stored · open-source, auditable gateway
base_url)Routing to cheaper models means some traffic leaves your environment. We don't hide that. You choose, per task, where data flows — and we put it in writing before any work begins. This is the part most providers won't show you.
Maximum savings (70–90%). Prompts are sent to the provider's servers. Best for non-sensitive, non-PII workloads.
When a task needs a top-tier model, or data is regulated, traffic stays on Western infrastructure you already trust.
DeepSeek / Qwen / Kimi weights run on servers you control. Nothing leaves your environment — zero egress.
| Route | Where data goes | Retention | Encryption | Keys |
|---|---|---|---|---|
| Cloud cheap | DeepSeek / Kimi / Qwen servers (Asia) | Provider policy — we store nothing | TLS 1.2+ in transit | Yours (BYOK) |
| Western flagship | OpenAI / Anthropic (US / EU) | Provider policy — we store nothing | TLS 1.2+ in transit | Yours (BYOK) |
| Self-hosted | Your infrastructure | Zero egress — you control | TLS + your KMS | Yours (BYOK) |
Non-sensitive / cost-driven: cloud cheap route saves the most. EU personal data or regulated: self-hosted or Western flagship — never the Asia route. US commercial: any route is fine. We document your chosen paths in writing before deployment — no surprises.
A self-hosted LiteLLM gateway sits between your app and the model providers. It routes each call by task type, applies quality gates, and falls back to a flagship model only when a task truly needs it. You point base_url at it — nothing else in your code changes.
Real numbers from the open-source CLI on a sample 50M-prompt / 20M-completion month at public 2026-07 list rates. Your mix will differ — the CLI measures yours, not a guess.
We map your real API calls, tag each task type, and show exactly where you're overpaying — a written report with numbers, not vibes.
A self-hosted LiteLLM gateway routes by task type, with quality gates and automatic fallback to a flagship model when a task truly needs it. You change base_url — nothing else in your code.
As models and prices shift, we re-route and re-tune so your bill stays low without you babysitting it. Quality gates and budget caps included.
Each call goes to the cheapest model that meets your quality bar — simple tasks on DeepSeek/Kimi, hard ones on GPT/Claude only when required.
Automatic fallback to a flagship model on low-confidence or failing responses, so output quality never silently drops.
Per-key max_budget hard limits. No surprise spikes, no open-ended exposure.
Bring your own DeepSeek / Kimi / Qwen / OpenAI / Anthropic keys. Your provider bills you; we never touch your keys.
Use your own monthly token counts. Prices are public list rates at time of writing (DeepSeek V4-Flash: $0.14/$0.28; Kimi K2.6: $0.95/$4.00; GPT-4o: $2.50/$10.00; GPT-4.1: $2/$8; Claude Sonnet 4.6: $3/$15; Claude Opus 4.8: $5/$25 per M). Verify current rates before any engagement.
Estimates from list prices; your real mix may differ. The open-source CLI does this from your actual logs — no guessing.
Solo-operator rates for SMBs. Prices climb toward market rates as we collect case studies. Payment via PayPal or bank transfer for early engagements.
We tell you exactly where your prompts go. Most providers won't.
Routing to cheaper models means some traffic leaves your environment. We put the data path in writing before any work begins — no surprises, no fine print.
Bring-your-own-keys. We never see your provider API keys. Your provider bills you directly.
Prompts aren't stored after delivery. No training on your data, ever.
Open-source, auditable. The gateway is LiteLLM (Apache-2.0). You can read every line we deploy.
Data residency is your choice. Route to self-hosted open-weight models (DeepSeek/Qwen/Kimi) on infrastructure you control — nothing leaves your environment.
You can — and many do. The paid service is for teams that lack the time to wire routing, tune quality gates, and monitor budgets monthly. The CLI is free precisely so you can validate the number first.
No. The gateway is OpenAI-compatible. You point base_url at it; the rest of your code is unchanged.
Not silently. Simple tasks move to cheaper models; hard tasks stay on a flagship via the quality gate plus fallback. You set the quality bar.
No. Even $100–500/mo bills often hide 50%+ waste on tasks that don't need a flagship. The free CLI tells you your number in under a minute.
Two options: (1) self-host open-weight models on your own infra, or (2) keep sensitive routes on Western flagship models. We document the data path per route before any work begins.
Yes, with one rule. China has no GDPR adequacy decision, so EU personal data (PII) shouldn't be sent to Asia-region model servers without a transfer mechanism. For those workloads, choose the self-hosted open-weight mode or keep them on Western flagship models. Non-personal or non-regulated traffic can use the cheap route. We document your chosen paths before deployment.
No. We never store your prompts after delivery, and we never train on customer data. The gateway runs on infrastructure you control; your provider bills you directly via your own keys. You can read every line — it's LiteLLM (Apache-2.0).
On the cloud cheap route, prompts are processed by DeepSeek / Kimi / Qwen on their servers — that's the trade-off for the low price, and we state it plainly rather than hiding it. If that's unacceptable for a given workload, route it to self-hosted or Western flagship instead. You decide per route; nothing is automatic.
Tell us about your current AI spend. We'll reply within 2 business days with a rough saving estimate — no cost, no obligation.
Prefer email? Write to [email protected]
Secure form · we reply within 2 business days.
Run the open-source CLI on your own token counts, or book a 20-minute audit. No card, no obligation.