OpenAI reportedly rolls out “Cyber” model to select users, signaling fresh rivalry with Mythos

New model, cautious rollout

It has been reported that OpenAI has begun offering a new model called “Cyber” to a limited set of users, a move that places the company more directly in competition with emerging rivals such as Mythos. The reported release is narrow and experimental. OpenAI appears to be testing capabilities and controls before any broad public launch. Why a small rollout? Because stakes are high — performance gains bring new product edges, but also new safety and regulatory risks.

Safety questions loom after clinical-style study

The announcement comes amid growing evidence that generative chatbots can be unreliable on health topics. A study published in BMJ Open found that across five leading AI platforms — ChatGPT, Gemini, Meta AI, Grok and DeepSeek — roughly half of responses to medical questions were “problematic,” and nearly 20% were judged “highly problematic.” The bots performed better on closed, factual questions and vaccine- or cancer-related prompts, and worse on open-ended topics such as stem-cell research or nutrition. It has been reported that OpenAI says more than 200 million weekly users ask ChatGPT health-related questions, and that the company intends to roll out specialized health tools for general users and clinicians — but the BMJ Open authors warn that these systems “can sound authoritative but be flawed.”

Competition, regulation and the trade-offs ahead

Reportedly, competitors including Anthropic have also been moving into healthcare-focused offerings, intensifying product rivalry even as regulators in the U.S., Europe and elsewhere tighten oversight of AI systems. The geopolitical backdrop matters: export controls, cross‑border data rules and national security concerns are shaping which models and datasets companies can access and where they can scale. Can careful, staged rollouts like Cyber’s balance innovation with public safety? Industry watchers say independent clinical validation, transparent evaluation and stronger guardrails will be required before these models are trusted for medical advice.