.@tryramp cut their AI costs by 30% with an LLM router.
Now they're giving it away for free.
Router started as internal tooling to keep 100+ AI features at Ramp on the right model. It routes 2.75 trillion tokens a month today, and they're opening it up to outside users.
What it does:
- One endpoint for OpenAI, Anthropic, Gemini and open-source models (incl. Kimi)
- Sends every request to the cheapest model that clears a quality bar
- Tests new models on real work each week, then re-routes automatically
- Layers caching, compression and 100+ optimizations on each request
- Falls back to another model when a provider goes down or rate-limits you
Using it is a one-line change: the API is OpenAI-compatible, so you swap your base URL and you're done. Free at launch (you pay list price for tokens), and the first 500 off the waitlist get $100 in credits.
Why now? Ramp processes AI vendor payments for thousands of companies. Their data shows business token usage up 1,001% from Jan 2025 to April 2026, with spend up 497%. @chamath says AI costs at his startup 8090 more than tripled since November and could hit $10m a year.
Ramp's mission is cutting costs. First they shipped a dashboard to see AI spend. Now the router to control it.
Model providers earn more when you use more frontier intelligence. "Ramp wins when you spend less."
This is exactly how a non-AI company becomes relevant in AI.
Solve the AI jobs to be done.
It makes you more newsworthy, more investor-friendly, and more useful to the fast-growing startups Ramp covets.
And here's a lovely detail. You don't need a Ramp account to use it.
This competes with OpenRouter, but doesn't have the same margin requirements, because it's also Ramp's top of funnel into the broader spend management platform.
Smart.
Now they're giving it away for free.
Router started as internal tooling to keep 100+ AI features at Ramp on the right model. It routes 2.75 trillion tokens a month today, and they're opening it up to outside users.
What it does:
- One endpoint for OpenAI, Anthropic, Gemini and open-source models (incl. Kimi)
- Sends every request to the cheapest model that clears a quality bar
- Tests new models on real work each week, then re-routes automatically
- Layers caching, compression and 100+ optimizations on each request
- Falls back to another model when a provider goes down or rate-limits you
Using it is a one-line change: the API is OpenAI-compatible, so you swap your base URL and you're done. Free at launch (you pay list price for tokens), and the first 500 off the waitlist get $100 in credits.
Why now? Ramp processes AI vendor payments for thousands of companies. Their data shows business token usage up 1,001% from Jan 2025 to April 2026, with spend up 497%. @chamath says AI costs at his startup 8090 more than tripled since November and could hit $10m a year.
Ramp's mission is cutting costs. First they shipped a dashboard to see AI spend. Now the router to control it.
Model providers earn more when you use more frontier intelligence. "Ramp wins when you spend less."
This is exactly how a non-AI company becomes relevant in AI.
Solve the AI jobs to be done.
It makes you more newsworthy, more investor-friendly, and more useful to the fast-growing startups Ramp covets.
And here's a lovely detail. You don't need a Ramp account to use it.
This competes with OpenRouter, but doesn't have the same margin requirements, because it's also Ramp's top of funnel into the broader spend management platform.
Smart.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
