gateway healthy · 99.97% 30d
Gateway 6.1 / 12 providers / 47 regions

Route any promptto the cheapest modelthat meets the bar.

Cortex is the orchestration plane for production model traffic. Versioned prompts, eval-gated promotions, sub-50ms failover, and a routing policy you can read.

Start free trialView docs
47ms
failover
94.6%
routing acc.
12
providers
policy / chat-prod
v 12 · signed
if intent == "summarise" and tokens < 4k:
→ haiku-4 · $0.0008
elif intent == "code":
→ sonnet-4.6 · $0.012
elif regulated:
→ self-hosted-llama · eu-only
else:
→ router.choose()
last 60s
0s421 req/s60s
182ms
p95
$0.0021
avg req
0.94
quality
Mercury OS▪Substrate▪Northwind Labs▪Foundry 47▪Plover Studio▪Helvetia Cartographic▪Field Notes Co.▪Linnea Bakery▪Marlow Workwear▪Verge Editions▪Mercury OS▪Substrate▪Northwind Labs▪Foundry 47▪Plover Studio▪Helvetia Cartographic▪Field Notes Co.▪Linnea Bakery▪Marlow Workwear▪Verge Editions▪
02 / Platform

Four surfaces. One control plane.

Router
haiku-4$0.0008 62%
sonnet-4.6$0.012 23%
opus-4.7$0.084 9%
self-hosted$0.0001 6%

A learned classifier over your traffic. Picks the cheapest model that hits your quality threshold, every request.

Latency p95
Providers
OpenAI●
Anthropic●
AWS●
Cohere●
Mistral●
Google●
Eval gate
PR
quality Δ
cost Δ
gate
summarise-v12 → v13
+0.04
-12%
pass
code-explain-v8 → v9
−0.02
+3%
fail
intent-classify-v4 → v5
+0.07
0%
pass
Cost ledger
this week
$8,427.13
vs last
−18.4%
haikusonnetopuslocalmisc
03 / In production

Three workloads. One gateway.

Cortex does not care if you serve a chat app, a batch pipeline, or a regulated workflow. The contract is the same: write the policy, ship the prompt, watch the meter.

01
SLA-bound product

Keep p95 under 800ms while shipping new prompts weekly.

Cortex routes around degraded providers in 47ms, falls back deterministically, and bills the difference to the right cost center automatically.

0
% reduction in p95 latency
02
Mixed-provider stack

Run cheap on hot paths, premium on the long tail.

Static rules and learned classifiers route requests to the lowest-cost model that meets the quality bar. The router is auditable. The savings are real.

0
% of traffic on tier-2 models
03
Regulated workloads

Prompts are code. Versioned, reviewed, signed.

Every change to a production prompt goes through pull request and gets a deterministic eval before it hits live traffic. Rollback is one click.

0
minute mean time to rollback
04 / Vs raw SDK

Your SDK. Civilised.

Capability
Cortex
Raw SDK
Provider failover
47ms automatic
Manual incident
Prompt versioning
Git-native, signed
Inline strings
Cost routing
Per-request, learned
Static config
Eval before promote
Required by default
Optional, skipped
Audit trail
Per-token, exportable
Aggregate logs
field reports

“We cut the model bill 41% in one quarter and reduced p95 by 320 milliseconds. Same quality. Same prompts. Better routing.”

Kohei Tanaka
Platform Lead, Mercury OS
0B
requests routed
0ms
median failover
0%
median cost saved
0%
gateway uptime
05 / Pricing

Pay for routing. Not for tokens.

Hobby
$0free forever

For prototypes and weekend builds.

  • 10k requests / mo
  • 3 prompt versions
  • Single environment
  • Community Discord
Start building
Team
$98per seat / month

For product teams shipping AI features.

  • 3M requests / mo
  • Unlimited versions
  • 3 environments + previews
  • Eval gating + diff
  • OTel + Datadog export
  • Business-hour support
Start 14-day trial
Enterprise
—annual contract

For regulated and high-volume teams.

  • Unlimited requests
  • Dedicated routing pool
  • VPC peering, BYOK
  • SCIM, audit log export
  • Solutions architect
  • 99.97% SLA
Contact sales
06 / FAQ

Direct answers to direct questions.

You write a routing policy. It can be static (always GPT for X), learned (classifier picks the cheapest model that scores above 0.86 on your eval), or composite. Every decision is logged with the inputs that produced it.

Recent changes

all releases
Apr 23
routing
Learned router supports per-customer overrides for enterprise tenants.
Apr 12
evals
Eval-on-PR posts a comment with cost delta, latency delta, and quality delta.
Apr 02
sdk
Python SDK 6.1 ships streaming with backpressure and clean cancellation.
Mar 21
regions
Tokyo gateway is generally available; ap-northeast-1 ready for traffic.

Replace your SDK in 8 lines.

Drop-in OpenAI and Anthropic compatibility. First 10,000 requests free, no card.

Start free Talk to sales