Are the Anonymous Horizon Models GPT-5 Early Release Models?

by

There is widespread speculation in the AI community that they represent a stealth preview, variant, or test version of OpenAI’s upcoming GPT-5 model. They have exceptional performance, timing, and shared characteristics with OpenAI’s tech stack.

The Horizon models (primarily Horizon Alpha and Horizon Beta) surfaced anonymously around July 30-31, 2025, on platforms like LM Arena and OpenRouter. They were noted for crushing benchmarks in coding, agentic tasks, and complex reasoning, often outperforming models like Claude 4 Sonnet.

Users and testers describe Horizon as the best code model ever with strengths in one-shot problem-solving, software development, and avoiding hallucinations.

Horizon followed a wave of anonymous models in LM Arena (e.g., Zenith, Summit, Lobster, Starfish, Nectarine, o3-alpha), which some believe were earlier GPT-5 components being assembled and tested in real-time. Zenith is also believed to be a GPT5 variant.

NOTE: The benchmarks are not telling the tale. The Claude models are still the best for real world coding even though they are behind other models on benchmarks.