The switch for AI inference.

Your provider keys. Your contracts. Every routing decision explained.

How it works

Three steps. No credits.

01
Connect your provider keys
Add the OpenAI, Anthropic, Gemini or OpenAI-compatible accounts you already pay for. Tandex doesn’t sell tokens.
02
Point your client at Tandex
Change the base URL in any OpenAI SDK and ask for tandex/auto.
03
See why each request went where
Each response can carry the route Tandex chose and the reason it chose it.

How it routes

One request in.
The right provider out.

Tandex measures the providers you connect and scores each request against those measurements. The scoring is deterministic: no model is asked to choose a model.

Time to first token

A first-class metric, measured per provider and model from real traffic, not read from a price sheet.

Cost

Set a cost ceiling per request. Routes above it are rejected; among the rest, cost is weighed alongside latency, load and reliability.

Region & residency

Limit each request to the providers and regions you approve, and to endpoints you have zero-retention terms with. Tandex enforces your choices; your contracts establish residency.

Availability

When a provider is slow or unavailable, new requests go to the next eligible route. Your code does not change.

Why Tandex

Not a proxy. A switch.

A proxy forwards your request. A switch decides where each request should go, records why, and keeps working when a route fails.

Decisions you can read

Every route is scored by rules, not by another model. Each response can carry why that provider was chosen.

Degrades, doesn't fall over

Written in Erlang/OTP, the runtime designed for telephone switches, where faults are contained and restarted rather than allowed to spread.

Your keys, your contracts

Bring your own provider keys. Tandex doesn't resell tokens, so your pricing, limits and data terms stay yours.

API

Change one line. Keep your SDK.

Point any OpenAI client at Tandex and ask for tandex/auto. Add the optional routing block when you need limits on latency, cost, region or retention.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.tandex.io/v1",
    api_key="TANDEX_API_KEY",
)

reply = client.chat.completions.create(
    model="tandex/auto",
    messages=[...],
    extra_body={
        "tandex": {
            "max_ttft_ms": 400,
            "region": "uk",
            "zero_data_retention": True,
        }
    },
)

Available to early-access teams. Routing fields shown are illustrative.

Early access

Route your first request through Tandex.

hello@tandex.io ↗

Tell us what you run, which providers you use and what latency you need. Early access is open to teams with real-time or regulated workloads.