Reliability
The failures that pass every health check
Five modes across two axes. The severe ones return HTTP 200 and corrupt state.
Case studyarXiv:2607.17525
Your AI stays online, even when providers don't.
One OpenAI-compatible endpoint for every model. Metriqual routes each call, fails over when a provider stalls, and logs all of it.
- OpenAI
- Claude
- Geminisoon
- Mistralsoon
- MiniMaxsoon
One key. Every model.
Change the base URL and the API key. Your OpenAI SDK code works exactly as it did.
-base_url="https://api.openai.com/v1"+base_url="https://api.metriqual.com/v1"-api_key=OPENAI_API_KEY+api_key=METRIQUAL_API_KEY
Not just chat.
Text, voice, video, image, music, image-to-video. One endpoint, one bill, one log.
Every call goes to the cheapest model that fits.
Priced live, per request. Routing costs under 10ms.
Then the provider stops answering.
The bad ones return 200 OK. They pass every health check and corrupt state quietly.
Reroute without losing the conversation.
99.2% of the conversation survives a mid-request failover. A stateless gateway keeps none of it.
744 of 750 injected failures · +59ms mean added latency · arXiv:2607.15899 · ContinuityBench
The response comes back. So does the receipt.
Every request logged and attributable. Cost, latency, model, agent.
All of that cost you two lines.
The same diff you saw at the start, unchanged.
-base_url="https://api.openai.com/v1"+base_url="https://api.metriqual.com/v1"-api_key=OPENAI_API_KEY+api_key=METRIQUAL_API_KEY
Every provider.
Every modality.
One key.
2 linesto integrate
<10msrouting overhead
99.2%conversation history carried
100%of requests logged & attributable
2 liveproviders — 3 more landing
Everything the gateway does, once you are through it.
Live web search
Answers from today, with the sources attached.
One flag and your agent reads the live web before it replies, then cites what it read. Works on OpenAI and Claude through the same endpoint.
GPT-5.1 pricing update[1]
openai.com
OpenAI cuts API prices again[2]
techcrunch.com
GPT-5.1 input pricing dropped ~20% this week[1]; output stayed flat[2].
Private knowledge
Your documents, answered from directly.
Upload what your team already wrote. It is chunked, embedded and retrieved per request, and none of it leaves your gateway on the way there.
"what is our refund window?"
Vector store2,172 chunks
top 4 retrievedyour vectors stay in your gateway
Predictable costs
A key cannot spend more than you let it.
Set a cap per key. Spend is attributed in real time across every request, provider and modality, so the month never ends in a surprise.
sk-live-prodthis month
$412/ $500 cap
sk-test-eval$200 / $200
cap reachednew requests rejected
PII detection
Customer data stops at your gateway.
Emails, cards and phone numbers are detected and masked in the request body before it is ever forwarded to an external provider.
Outbound requestraw
refund for jane@acme.io on card 4242 4242 4242 4242, phoned from +1 415 555 0132
3 entities masked before the provider saw them
Prompt hub
Edit the prompt without shipping the app.
Version prompts, see which one is live, and point any version at any provider. Changing a system prompt stops being a deploy.
support-replyv2 → v3
deployed You are a support agent.
−Decline refunds politely.
+Decline in one line, link the refund policy.
+Always cite the doc you used.
Keep replies under three sentences.
GPT-5.1 pricing update[1]
openai.com
OpenAI cuts API prices again[2]
techcrunch.com
GPT-5.1 input pricing dropped ~20% this week[1]; output stayed flat[2].
Live web search
Answers from today, with the sources attached.
One flag and your agent reads the live web before it replies, then cites what it read. Works on OpenAI and Claude through the same endpoint.
"what is our refund window?"
Vector store2,172 chunks
top 4 retrievedyour vectors stay in your gateway
Private knowledge
Your documents, answered from directly.
Upload what your team already wrote. It is chunked, embedded and retrieved per request, and none of it leaves your gateway on the way there.
sk-live-prodthis month
$412/ $500 cap
sk-test-eval$200 / $200
cap reachednew requests rejected
Predictable costs
A key cannot spend more than you let it.
Set a cap per key. Spend is attributed in real time across every request, provider and modality, so the month never ends in a surprise.
Outbound requestraw
refund for jane@acme.io on card 4242 4242 4242 4242, phoned from +1 415 555 0132
3 entities masked before the provider saw them
PII detection
Customer data stops at your gateway.
Emails, cards and phone numbers are detected and masked in the request body before it is ever forwarded to an external provider.
support-replyv2 → v3
deployed You are a support agent.
−Decline refunds politely.
+Decline in one line, link the refund policy.
+Always cite the doc you used.
Keep replies under three sentences.
Prompt hub
Edit the prompt without shipping the app.
Version prompts, see which one is live, and point any version at any provider. Changing a system prompt stops being a deploy.
< Change two lines.
Keep everything else. >
Swap your API key and base URL for Metriqual's. Everything else in your codebase, from prompts to parameters, stays exactly as it is.
metriqual — live trafficLIVE
YOUR ENTIRE MIGRATION
- api_key = "OPENAI_API_KEY" + api_key = "MQ_KEY" - base_url = "https://api.openai.com/v1" + base_url = "https://api.metriqual.com/v1"
WHAT HAPPENS WHEN A PROVIDER FAILS
$ POST /v1/chat/completions model=gpt-4o
→ openai connect 2.00s … timeout
↻ failover next step in chain
→ anthropic claude-fable-5 · 200 OK · 412ms
✓ delivered 0 requests dropped
Everything a gateway should do, in one place.
 Metriqual | Direct provider | Other gateways | |
|---|---|---|---|
| Automatic failover across providers | |||
| 2-line integration (no SDK rewrite) | |||
| Conversation history carried on failover | |||
| Prompt versioning & hub | |||
| Per-agent spend controls | |||
| Multi-modal (voice, video, image, music) | |||
| Fully managed, nothing to host | |||
| Unified observability |
Swipe to compare providers→
| Metriqual | Direct provider | Other gateways | |
|---|---|---|---|
| Automatic failover across providers | |||
| 2-line integration (no SDK rewrite) | |||
| Conversation history carried on failover | |||
| Prompt versioning & hub | |||
| Per-agent spend controls | |||
| Multi-modal (voice, video, image, music) | |||
| Fully managed, nothing to host | |||
| Unified observability |
partial / limited
Zero dropped requests, even mid-outage.
Failover reroutes every request to a healthy provider in milliseconds. A single-provider setup goes down with the provider.
Metriqual Failover active
0failed requests
The request is replayed against the next provider in the chain, carrying its conversation history. Your user sees a slower answer, never an error.
Failover overhead
<10ms
Providers in chain
5
History carried
Yes
Single-provider setup Provider down
100%of requests fail
While the provider is down your users get an error page, not an answer. Retries hit the same endpoint, so the outage lasts as long as theirs does.
Failover
None
Providers
1
History carried
No
We measured it, then published the method.
Failover that keeps the request alive but loses the conversation isn't failover. We built the benchmark for it and ran it on our own gateway.
ReliabilityarXiv:2607.17525
The failures that pass every health check
Five modes across two axes. The severe ones return HTTP 200 and corrupt state.
- Five failure modes mapped across detectability and origin.
- The severe ones answer 200 OK and quietly corrupt state.
- Health checks stay green while the answers rot.
More
Built for every stage.
01
Free
EVALUATE & PROTOTYPE$0
- 1 agent endpoint
- 10K requests / month
- 3 providers
- No automatic failover
- Community support
02
Builder
SHIP YOUR FIRST AI FEATURE$0/month
- 5 agent endpoints
- 100K requests / month
- All 5 providers
- Automatic failover
- Prompt Hub + versioning
- 30-day logs
- Email support
RECOMMENDED
Growth
PRODUCTION AT REAL USAGE$0/month
- Unlimited agent endpoints
- 500K requests / month
- Per-agent spend controls
- PII redaction
- Webhooks + A/B routing
- Team access (5 seats)
- Priority support
04
Scale
AI AS CORE INFRASTRUCTURE$0/month
- Unlimited everything
- 2M+ requests / month
- Conversation history SLA
- Custom fallback logic
- Datadog/Splunk export
- 1-year retention
- Dedicated Slack
05
Enterprise
DOWNTIME = REVENUE IMPACTCustom
- Everything in Scale
- Custom SLAs with penalties
- Dedicated solution engineer
- Security & compliance review
- Annual contract
- Co-authored case study
- Unlimited agent endpoints
- 500K requests / month
- Per-agent spend controls
- PII redaction
- Webhooks + A/B routing
- Team access (5 seats)
- Priority support
Frequently asked questions.
Everything developers need to know before building production-ready AI agents.
Three steps to an agent that stays up.
- Agent nameRESEARCH_AGENTModelopenai · gpt-4o01
Name it, point it at a model
Pick a provider and a model. That is the whole configuration.
- openai · gpt-4oprimaryanthropic · claude-sonnet-4-5fallback02
Give it a fallback
Add a second provider. If the first one stalls, traffic moves there automatically.
base_url = "api.metriqual.com/v1" api_key = MQ_KEY
Live, with failover armed03Ship it
Swap your base URL. Every request now routes through the chain you just built.
Watch a provider go down.
Thirty minutes with an engineer. We kill a provider mid-conversation and you see exactly what your users would have seen.
✓ Production ReadyAutomatic failoverSub-10ms Overhead