OpenRouter+10 Mixed

Operator Insights

What are teams learning from building on OpenRouter and running it in production?

5 accounts: 1 failure, incident or rollback, 1 left it or turned it down and 3 adopted it

Failures, incidents and rollbacks1

Mo MoustafaFounderOlly
Olly, an AI assistant in iMessage that has carried over 18 million messages, about a third on open models through OpenRouter.
mmoustafa.com, published Sep 7, 2026Read the write-up (opens mmoustafa.com in a new tab)
Observation
OpenRouter's per-host benchmarks, reported by Olly's founder, scored the same DeepSeek weights 90% on GPQA at DeepSeek and 75% at another host, and Olly went down after pinning three hosts with fallbacks off.
Takeaway
Choose hosts by the benchmark closest to your workload and keep fallbacks on, because a hard filter also shrinks the pool OpenRouter can fall back to.
What happened
Olly routes open models such as DeepSeek V4 Flash through OpenRouter. OpenRouter's per-host board showed first-party DeepSeek at 90% GPQA and 81% TAU-Bench Airline, against 75% and 58% for DigitalOcean on the same weights, and in July one host carried about 20% of Olly's DeepSeek traffic and 92% of its empty completions. Filtering hosts by declared precision did not predict quality. Pinning the model to three hosts with fallbacks off took Olly down within two weeks as those hosts rate limited or stopped serving it.
- Public engineering write-up
- Measured results reported
Deployment details
Olly, an AI assistant in iMessage that has carried over 18 million messages, about a third on open models through OpenRouter.
- Public engineering write-up
- Measured results reported
Team
Olly, run by its founder
Workload
Over 18 million assistant messages, about a third on open models through OpenRouter
Model
DeepSeek V4 Flash 0731 and other open models
Measurement
GPQA 90% and TAU 81% first-party against 75% and 58% at DigitalOcean (Sep 7, 2026); 92% of empty completions from one host carrying about 20% of traffic (July 2026)

Left it or turned it down1

Jay BrunetSystem Admin and Developer at SRVpress, by his X profileSRVpress
SRVpress, a small hosting and development shop, replying on X to another developer who had left OpenRouter.
X, published Oct 25, 2026Read the post (opens X in a new tab)
Observation
SRVpress stopped using OpenRouter over slowness and errors, the same complaint the developer it answered had raised.
Takeaway
Track a router's latency and error rate per model and host from your own logs, and keep a direct provider key ready, because the teams that leave name speed and errors.
What happened
On October 25, 2026, SRVpress's system admin and developer wrote on X that he had stopped using OpenRouter for slowness and errors, agreeing with a developer who gave the same reason. The post gives no traffic volume, model or replacement.
“Stopped using OpenRouter for same reason, slowness & errors.”
Jay Brunet
- Public practitioner write-up (X post)
- No metrics reported
Deployment details
SRVpress, a small hosting and development shop, replying on X to another developer who had left OpenRouter.
- Public practitioner write-up (X post)
- No metrics reported
Team
SRVpress

Adopted it3

Aleks MistratovHead of AI ProductDescript
Descript's video editing agent, Underlord, which runs on frontier models from Anthropic, Google, OpenAI, and xAI.
OpenRouter blog, published Sep 15, 2026Read the case study (opens OpenRouter blog in a new tab)
Observation
Descript moved its Underlord editing agent from three direct model integrations to one OpenRouter connection and runs 13 models from four developers in production.
Takeaway
Consolidate model access when evaluating new models slows you down more than integrating them. Set host order and data policy per model and choose each route yourself.
What happened
Descript maintained separate integrations with OpenAI, Anthropic, and Google and wrote its own fallback logic between them, so each new model waited for an engineer. With OpenRouter as its production inference layer, it sets host order per model from OpenRouter's throughput, uptime, and cost data, routes only to hosts within its data retention policy, and configures 3 to 4 automated fallbacks per model. It keeps a direct Anthropic connection outside production for early-access models.
“Production went from three direct integrations to one connection through OpenRouter.”
OpenRouter
- Vendor customer story
- Team metrics reported
Deployment details
Descript's video editing agent, Underlord, which runs on frontier models from Anthropic, Google, OpenAI, and xAI.
- Vendor customer story
- Team metrics reported
Team
Descript AI product (Underlord agent)
Workload
Underlord, Descript's video editing agent
Model
13 models from Anthropic, Google, OpenAI, and xAI as of August 2026
Measurement
New-model evaluation from 1 week or more to 1 to 2 hours; 3 to 4 automated fallbacks per model

Kilo CodeJun 5, 2025Kilo BlogKilo Code's coding-agent traffic runs through OpenRouter, and three months after launch the agent reached third place on OpenRouter's weekly app rankings.
Justin HalsallFirst hireKilo Code
Kilo Code, an open-source AI coding agent whose users reach models through OpenRouter.
Kilo Blog, published Jun 5, 2025Read the post (opens Kilo Blog in a new tab)
Observation
Kilo Code's coding-agent traffic runs through OpenRouter, and three months after launch the agent reached third place on OpenRouter's weekly app rankings.
Takeaway
A tool that hands users model credits needs abuse controls from the first day, because free credits on a router attract account farming.
What happened
Kilo Code launched in March 2025 with a ten-person team assembled within a week. In its three-month review, its first hire lists the agent at third place on OpenRouter for the week and sixth for the month among its milestones, and says Kilo Code takes no cut of token usage, so it has no reason to push consumption. The same post reports a tool built to harvest the free credits Kilo Code gave new users by registering tens of thousands of accounts.
Public organization write-up
Rankings reported
Deployment details
Kilo Code, an open-source AI coding agent whose users reach models through OpenRouter.
Public organization write-up
Rankings reported
Team
Kilo Code, a ten-person team
Duration
Three months, March to June 2025
Workload
An open-source coding agent's model traffic
Measurement
Third on OpenRouter's weekly app rankings and sixth for the month (Jun 5, 2025)
Maksim DanilchenkoOct 5, 2026DEV Communitysoba routes every meal-photo recognition call through OpenRouter with strict JSON Schema and a model fallback list, at US$0.005 to US$0.01 per scan in production.
Maksim DanilchenkoBuilder of sobasoba
soba, an iOS app for people with type 1 diabetes that estimates carbs from a photo of a plate.
DEV Community, published Oct 5, 2026Read the write-up (opens DEV Community in a new tab)
Observation
soba routes every meal-photo recognition call through OpenRouter with strict JSON Schema and a model fallback list, at US$0.005 to US$0.01 per scan in production.
Takeaway
Use the router to swap models against a benchmark and enforce one output schema, and measure cost and latency per request in your own billing and logs.
What happened
soba sends each scan through OpenRouter and forces strict JSON Schema output there. Its builder picked the model with a published food-recognition benchmark, then moved from Gemini 3.1 Pro to gemini-3.6-flash, which cut cost per scan from about US$0.012 to about US$0.005 and latency from 7.2 to 8.3 seconds to 2.6 to 6 seconds, with quality differences inside the noise of gram estimates. The write-up credits strict schemas and a prompt built around scene scale with most of the accuracy gain.
Public practitioner write-up
Production cost and latency reported
Deployment details
soba, an iOS app for people with type 1 diabetes that estimates carbs from a photo of a plate.
Public practitioner write-up
Production cost and latency reported
Team
soba, built by Maksim Danilchenko
Workload
Meal-photo recognition for an iOS carb-counting app
Model
gemini-3.6-flash, after Gemini 3.1 Pro
Measurement
US$0.005 to US$0.01 per scan and 2.6 to 6 seconds per scan in production (Aug 29, 2026)

Failures, incidents and rollbacks1

OpenRouter's per-host benchmarks, reported by Olly's founder, scored the same DeepSeek weights 90% on GPQA at DeepSeek and 75% at another host, and Olly went down after pinning three hosts with fallbacks off.

Left it or turned it down1

SRVpress stopped using OpenRouter over slowness and errors, the same complaint the developer it answered had raised.

Adopted it3

Descript moved its Underlord editing agent from three direct model integrations to one OpenRouter connection and runs 13 models from four developers in production.

Kilo Code's coding-agent traffic runs through OpenRouter, and three months after launch the agent reached third place on OpenRouter's weekly app rankings.

soba routes every meal-photo recognition call through OpenRouter with strict JSON Schema and a model fallback list, at US$0.005 to US$0.01 per scan in production.