OpenAI launches GPT-6 models to all ChatGPT users globally

OpenAI has released GPT-6 Sol and GPT-6 Luna to all ChatGPT users across free and paid plans globally, replacing the previous GPT-5.6 Sol and GPT-5.6 Luna models in ChatGPT. The October versions described here are distinct from earlier September versions that remain in use in Codex and ChatGPT Work, according to a system card published by the company.

The new models incorporate safety advances from OpenAI's Astra framework and updated safety training targeting high-risk misuse in cybersecurity, biology, and violence. Under OpenAI's Preparedness Framework, both models are classified as "High" capability in Cybersecurity and Biological and Chemical domains, but neither reaches the "High" threshold for AI Self-Improvement.

Compared with their GPT-5.6 counterparts, the October models showed stronger resistance to multi-turn jailbreak attempts and reductions in deceptive behavior. However, evaluations found statistically significant regressions in self-harm, gore, and sexual content safety scores for GPT-6 Luna, and in self-harm scores for GPT-6 Sol. OpenAI said manual review of these failures found violations were generally low severity.

For users under 18, both models showed statistically significant regressions in age-restricted content, sexual content, and emotional reliance categories compared with the GPT-5.6 August models. OpenAI said it applies additional classifier-based filters for younger users that are not reflected in the published evaluation scores.

On health benchmarks, GPT-6 Sol and GPT-6 Luna outperformed their GPT-5.6 counterparts on HealthBench Professional, while showing slight regressions on standard HealthBench scores, which OpenAI attributed largely to a response-length scoring adjustment. Both models showed improvements on MentalHealthBench across all acuity levels.

On factuality, both models delivered improvements over their predecessors across most metrics covering user-flagged failures and high-stakes medical, legal, and financial prompts.

On cybersecurity benchmarks, GPT-6 Sol scored 82.62% on ExploitBench and 68.85% on SEC-Bench Pro, while GPT-6 Luna scored 44.66% and 48.50%, respectively.

You May Also Be Interested In

• Apollo Commercial Real Estate Finance to delist from NYSE by Nov. 6
• GreenPower receives Nasdaq notice for falling below $1 minimum bid
• Rubico prices $8.5M convertible note offering on Nasdaq