OpenRouter has quietly released a new large-language model dubbed “Horizon Alpha,” describing it only as a “stealth model” and inviting users to test it for free. Early users report the system supports a 256,000-token context window—four times the length offered by OpenAI’s flagship GPT-4o—and say it tops the community-run EQ-Bench leaderboard that evaluates reasoning and writing quality. The launch has fuelled industry speculation that Horizon Alpha may in fact be an unreleased version of GPT-5. Several independent developers pointed to the model’s performance and creative-writing abilities as evidence, although OpenAI has not commented and no formal attribution appears in OpenRouter’s documentation. Horizon Alpha joins a growing roster of unannounced models—including “Zenith,” “Lobster” and “Summit”—that researchers have discovered on the routing platform in recent months. For now, the tool remains freely accessible, giving outside engineers an early opportunity to probe capabilities that could foreshadow the next generation of commercial generative-AI systems.