Relayv0.2
Sign inStart free

One API key. Unlocks different model capabilities.

Image gen, video gen, OCR, speech-to-text, text-to-speech, reasoning, code, tool calls — one key for every capability your agent needs.

Start free$pip install relay-ai-sdk
from openai import OpenAI

# One Relay key. Every model.
client = OpenAI(
    api_key="rly_...",
    base_url="https://api.relay.ai5labs.com/v1",
)

# Reasoning
client.chat.completions.create(
    model="claude-opus-4.8",
    messages=[{"role": "user", "content": "..."}],
)

# Code
client.chat.completions.create(
    model="claude-sonnet-4.6",
    messages=[{"role": "user", "content": "..."}],
)

# Fast & cheap
client.chat.completions.create(
    model="deepseek-v3",
    messages=[{"role": "user", "content": "..."}],
)
claude-opus-4.6claude-sonnet-4.6claude-haiku-4.5gpt-4ogpt-4o-minigemini-3-progemini-3.5-flashllama-3.3-70bllama-4-scoutqwen3-32bgpt-oss-120bdeepseek-v4-proglm-5.1kimi-k2.6mistral-largecommand-agrok-3and moreclaude-opus-4.6claude-sonnet-4.6claude-haiku-4.5gpt-4ogpt-4o-minigemini-3-progemini-3.5-flashllama-3.3-70bllama-4-scoutqwen3-32bgpt-oss-120bdeepseek-v4-proglm-5.1kimi-k2.6mistral-largecommand-agrok-3and more
Agentic workflows

Find the right model for every step of your workflow.

Reasoning
claude-opus-4.8
thinking · done

Okay — the user wants a refund on order #4523, the WH-1000XM6 headphones at $1,240. First, eligibility. The order shipped 6 days ago and was delivered yesterday. Our policy is 30 days from delivery, so they're comfortably within window. No restocking fee on a first return — that's clean. Pulling the call transcript: they said "the headphones never sounded right". Vague, but consistent with a defect claim, not buyer's remorse. No need to push back or offer a partial. Account history: signed up over a year ago. No prior refund flags, no chargebacks, no suspicious patterns. Three previous orders, all kept. Real customer, not someone gaming the policy. Cost check: $1,240 is under the $2,000 manager-approval threshold. I can process it directly without escalation. So: full refund to the original card, processed today. Email should be short and empathetic — acknowledge the disappointment, confirm the refund timing (5–7 business days), and offer a no-questions return shipping label. Don't speculate about what went wrong; let support follow up if they want a replacement.

Code
claude-sonnet-4.6
auth.ts · claude-sonnet-4.6+3 −1
12− const token = req.headers.authorization?.slice(7);
13+ const token = parseBearer(req.headers.authorization);
14 if (!token) return res.status(401).json({ error: 'no_token' });
15
16+ const { sub, exp } = await verifyJWT(token, JWKS);
17+ if (exp * 1000 < Date.now()) return res.status(401).json({ error: 'expired' });
18 req.userId = sub;
19 next();
Tool calls
gpt-4o
function call · strict200 OK
tool
github.create_issue
{
  "repo":   "ai5labs/relay-sdk",
  "title":  "Add retry policy for 429s",
  "labels": ["robustness", "p1"],
  "body":   "Provider 429s currently bubble..."
}
→ issue #427 created
Vision / OCR
gpt-4o
SOUNDFORM INC.
INVOICE
123 Brannan St · San Francisco, CA
Invoice #45232026-06-04
WH-1000XM6 Headphones$1,099.00
Carrying case$ 42.00
Shipping$ 18.00
Subtotal$1,159.00
Tax$ 81.00
Total$1,240.00
extracted
vendorSOUNDFORM INC.
invoice#4523
date2026-06-04
subtotal$1,159.00
tax$81.00
total$1,240.00
statusPAID
Image gen
imagen-4
snow-capped peaks at golden hour, dramatic clouds
1 / 4
prompt
snow-capped peaks at golden hour, dramatic clouds
Audio
whisper-large-v3
transcribing · 0:14 / 1:42● rec
[00:08] Caller  “Hey, I'm calling about order 4523 — the headphones never sounded right.”
[00:14] Agent  “Sorry to hear that. Let me pull it up — one moment.”
Video gen
veo-3
● generating
1280×720 · 6s
prompt
aerial drone shot · ocean waves at sunset, slow pan
High volume
deepseek-v3
inbox · triage1,847 / 14,050
  • priorityboard@jpmorgan.com · Q4 board agenda — please review4m
  • supportsupport@soundform.com · Refund stuck — order #45235m
  • spamnoreply@sweepstakes.io · You won $5,000,000!!! CLAIM NOW7m
  • prioritycalendar@notion.so · Investor sync — confirm 3pm slot9m
  • supportfeedback@soundform.com · App keeps crashing on iOS 2612m
  • archivenews@import-ai.com · Import AI #427 — top papers18m
Built in

Everything you need in production.

Real features that ship with v0.2 — not roadmap promises.

Cost tracking

Every response carries a cost and its source — never wrong by default.

Streaming

Tool deltas keyed by index, thinking blocks preserved.

Structured output

Pass a Pydantic model, get a validated object back.

MCP tool layer

Any MCP server, any model, one schema compiler.

Batch

Batch multiple requests behind one call, ~50% off.

Fallback groups

Retry on a backup model when the primary fails.

Pricing

Library is free. Gateway is pay-as-you-go.

Run it yourself with your own keys, or use the hosted Gateway for one bill across every model.

Open source
$0
Self-hosted, BYOK
  • – Every model and feature
  • – Keys stay in your process
  • – Snapshot pricing
  • – Community support
Read the docs
Hosted Gateway
Usage-based
pay only for what you use
  • – One endpoint, one bill
  • – One key for every model
  • – Usage, logs and latency
  • – Spend caps
Start free
Enterprise
v0.3 · Q3 2026
Talk to us
On the roadmap
  • – SLAs, VPC deployment
  • – SOC 2, BAA / DPA
  • – India-hosted inference
  • – Design partners welcome now
Enterprise & India
Partner with us

Build the inference layer together.

Become a design partner

First-mover pricing and a direct engineering channel as we build out the inference tier.

engineering@ai5labs.com
Invest or partner strategically

Preference for institutional funds with a sovereign-AI thesis, or strategic capital aligned with the IndiaAI Mission's compute deployment.

bryan@ai5labs.com