The world’s fastest model for computer use agents
Try LightconeIntroducing our fastest model yet, Lightcone 1.Computer use agents optimized for speed and accuracy. Over 92% grounding accuracy, at sub‑second latency.
Explore Lightcone's full suite of capabilities through our chat interface.
The more it runs, the less you touch.
Up to 0%
of tasks are handled
autonomously
As measured on internal benchmarks
In Unreal Engine, model a low-poly desert scene and render it at 1080p and make it animate
Unreal EngineFind the cheapest nonstop SFO→Tokyo over the next 3 weeks
BrowserScan through my email and find all the latest receipts from this year
GmailBuild a DoorDash grocery cart for a week of dinners under $120
DoorDashExtract patient info and triage based on urgency, and create a recurring priority list
BrowserSoftware was built for people. Now it can run itself.Describe the outcome and point Lightcone‑1 at any screen. It sees, decides, and acts — over and over until the job is done. No brittle scripts, no per‑site selectors.
Trained for agency
Our series of large language models, starting from Northstar and through our flagship Lightcone‑1 model, are trained from the ground up for agentic action taking – in addition to coding, tool use and being a helpful assistant.
Crafted with care (and Rust)
We care about speed and performance, so that you can work at the speed of light. With our custom inference engine built in Rust, we’re not here to let you wait.
Always ready, always working
We believe the future of work is collaborative–where humans focus on high-value decisions, while technology handles the rest, at scale. Our models run wherever you need – especially in the background.
Fast and accurate. Not a tradeoff.Most computer-use models make you choose. Lightcone‑1 sits in the corner everyone wants.
Over
0%
Grounding accuracy on ScreenSpot‑v2
Up to
0×
Faster than frontier computer use-models
Used by
0+
Developer teams
Median
0ms
Per action latency
Positions are illustrative. Latency from a 2026‑06 head‑to‑head computer‑use benchmark; accuracy reflects Lightcone‑1’s ScreenSpot‑v2 result (92.8%).