OpenAI-compatible inference routing

Inference routing that proves every decision.

Neural Router routes each AI request across models and providers using your cost, quality, latency, budget, and policy rules, then hands you the receipt to verify exactly what happened.

// live routing · 12 modelsPOLICY: QUALITY
requestGPT-5.4$$$Claude Opus 4$$$Gemini 3 Pro$$Grok 4$$Llama 4$DeepSeek 4$Qwen 3$Kimi K2.7$MiniMax 2.7$Mistral Large$Gemma 3$Command R+$
scored on 4 signalsClaude Opus 4
One OpenAI-compatible endpoint
Per-request route receipts
Model & provider controls
Exact usage & cost accounting
Why teams switch

Stop treating model routing like a black box.

Most gateways forward requests and fall back when something fails. That's table stakes. Production teams need to know why a request went to a model, whether it respected the policy, what it cost, and whether a cheaper path would have preserved quality. Neural Router turns routing into an auditable control plane where every decision is constrained, scored, traced, and explainable.

Route by outcome, not vibes

Match each request to the cheapest eligible model that satisfies the workload's quality, latency, capability, and governance requirements.

Show the work

Inspect the candidates, scores, constraints, cache state, fallback path, and final reason on every routing receipt.

Keep spend under control

Set budgets, service types, quality floors, and allowed models. Degrade gracefully before a runaway agent drains the account.

Hold providers accountable

Route only to providers that clear measured conformance, uptime, latency, reliability, and quality thresholds.

Verifiable routing

A receipt for every decision.

A routing receipt turns auto into an inspectable system. See the candidates considered, the constraints applied, the score tradeoffs, the model selected, and the cost, latency, and quality reason behind the route. No silent fallback. No prompt mutation.

  • Candidate models, scores, and exclusion reasons
  • Route reason + trace ID on every request
  • Token, cost, and latency that reconcile back to the call
  • Policy version on every route, so decisions are replayable
routing receipt
tr_a17f3
POST/v1/chat.completionsservice: precision

intent · contract clause extraction · json_schema

Claude Opus 4$$$

meets quality floor · lowest eligible cost

GPT-5.4$$$
Llama 4 405B$

below quality floor

Qwen 3$

no json_schema support

route reason

precision floor · quality ≥ 0.9

cost

$0.0142

ttft

410 ms

18% cheaper than the default model at the same quality floor
How it works

One endpoint. A real routing decision behind every call.

01

Connect your SDK

Point your OpenAI-compatible client at Neural Router and create a scoped API key for each workload.

02

Declare intent

Choose a service type such as Standard, Saver, Turbo, Precision, Scale, or Agent, or define your own routing policy.

03

Let the router optimize

Neural Router profiles the request, applies policy, checks provider health and capabilities, scores candidates, and selects the best eligible route.

04

Verify the receipt

Inspect the selected model, route reason, cost, latency, cache result, and policy version after the request.

route.py
from openai import OpenAI
client = OpenAI(base_url="https://api.neuralrouter.ai/v1",api_key="nr-••••",)
# declare intent and the router compiles the policyr = client.chat.completions.create(model="precision",messages=[{"role": "user", "content": "…"}],)
# inspect the route reason on the receiptr.headers["x-nr-route-reason"]# → "precision floor · quality ≥ 0.9"
Product pillars

Built for teams that measure inference like infrastructure.

Routing Studio

Create policies by objective, service type, allowed models, fallback order, budget, and quality floor. Use auto, cheapest, fastest, or pinned models without changing SDKs.

Trust Dashboard

See route receipts, token and cost reconciliation, provider health, conformance coverage, and router overhead in one place.

Provider Scorecards

Use active probes and passive telemetry to route around provider drift, elevated error rates, latency spikes, and capability gaps.

Enterprise Controls

Manage teams, workspaces, API keys, BYOK, audit logs, privacy settings, and observability exports from one governed surface.

Service types

Choose what the key is for. The router compiles the policy.

You shouldn't have to tune routing weights by hand for every workload. Neural Router packages common inference patterns into service types that map to concrete router constraints.

Standard

Balanced quality per dollar for everyday production traffic.

Saver

Cheapest route that clears a sane quality bar.

Turbo

Low time-to-first-token for interactive products.

Precision

Highest-quality route with strict quality floors.

Scale

Batch and large-volume workloads optimized for unit cost.

Agent

Multi-step workloads with budget, session, and cache affinity.

Custom

Full control over dimensions, constraints, and saved presets.

Compare service typesSee how each maps to pricing
Provider market

Better providers should win more traffic.

Neural Router evaluates providers on the metrics buyers actually care about: uptime, latency, token accounting, API conformance, capability support, quality drift, price, and reliability. Providers get transparent feedback and yield analytics; consumers get a cleaner supply pool.

List your endpoint

Conformance

Checked before traffic

Scorecards

Measured, not self-reported

Shadow audition

Prove quality before go-live

Yield analytics

See why you won or lost

Make your next inference route inspectable.

Start with one key, one workload, and one measurable goal.