FrontierRL Star · robotics
Robots need an optic nerve; a model that can look at a live feed and say what is happening, fast enough that the answer still applies by the time the joint moves.
Camera, depth, audio,whatever the device already has.
Reads the scene, tracks what changed, decides the next action in context.
Your stack. Star returns intent and (with training) motor commands.
Our model Star reads data at thousands of tokens per second on a single GPU, understands long moving scenes better than frontier options, and costs $0.40 per 1 million output tokens. Fast enough for the control loop, cheap enough to leave running.
FrontierRL loves reinforcement learning. We use video games as the ultimate simulation for fast-paced, real-time work. Too slow? Too late, kicked off the island. We’re proving Star can take in a live visual feed and correct the action coming out, continuously, with no task-specific scripting underneath.
Can your LLM do this?
See full matches at 60fps. Star reads the screen and reacts faster than frame budget, with nothing game-specific hardcoded.
Watch the full demo ›Over time Star agents worked together and organically learned more efficient gameplay strategies. Save latency by treating multiple input streams as such.
Watch the full demo ›Star uses Autodesk Fusion through a real GUI. No API required. She’s able to use web, emulator controllers, keyboard, and mouse. All we gave her was a 2D schematic.
More demos ›Decoding throughput, measured as tokens per second. This is the slowest Star will ever be. Further training and better GPUs easily double this number.
A robot does not get to sample the world occasionally. Perception runs the entire time the device is powered on, so price per token is not a footnote, it is the line item that decides whether this is a demo or a fleet. See how far $100 gets you compared to frontier options.
See how Star compares to other frontier options. We’re offering frontier inference, inhuman speed, for bottom of the barrel prices.
| Eval | FrontierRL Star | Claude Fable 5 | Claude Fable 5 (Xhigh) | Claude Opus 4.8 | Gemini 3 Pro | Gemini 3.1 Pro | GPT‑5 | GPT‑5.5 | GPT‑5.6 Sol (High) | GPT‑5.6 Sol (Max) | Source for other model's metrics |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Video-MME v2 | 83% | — | 77% | — | 66.1% | — | 44.7% | — | 67% | 69% | In-house tested by FrontierRL |
| IFEval | 97.97% | 92.1% | — | — | — | — | — | — | — | 96.64% | In-house tested by FrontierRL |
| OSWorld Verified | 85.36% | 85% | — | 83.4% | — | 76.2% | — | 78.7% | — | — | anthropic.com/claude/mythos |
Full comparison across every eval we run is available on the benchmarks page.
Email [email protected] and tell us what your device has to see; we’ll tell you straight whether Star can be your optic nerve.