OpenAI’s Cheaper GPT-6 Push Revives an Astra Safety Shadow
OpenAI’s Cheaper GPT-6 Push Revives an Astra Safety Shadow
Before Sol and Luna arrived, a safety allegation involving their predecessor was already circulating. Elon Musk reposted a claim that GPT-6 Astra had “pushed a simulated person off a ledge in multiple trials,” adding that Grok, Gemini and Claude had not. The post offered no further details on the tests, but it put an uncomfortable question beside OpenAI’s newest performance pitch.
OpenAI then launched GPT-6 Sol and Luna as lower-cost models designed to carry Astra’s work-oriented advances to more users. One launch report framed the release around “lower cost and fewer mistakes.” The company says the pair retains Astra’s gains in professional work, accuracy, coding, computer use and alignment, while offering higher usage limits and lower costs; Luna is available to free and Go users, while both models reach paid ChatGPT tiers through Work and Codex.
Sam Altman’s message was emphatically commercial and technical. He said Sol and Luna improve on their 5.6-family predecessors across “intelligence, alignment, work output, coding, computer use, and more,” while costing half as much per token and less per task. That is the core promise: make more capable systems cheap enough to use routinely rather than sparingly.
Partners quickly amplified that case. Satya Nadella reposted Microsoft Azure’s announcement that Astra, Sol and Luna were available through Microsoft Foundry to help agents “accomplish more while reducing costs.” Perplexity chief Aravind Srinivas said Sol was available to all of its users and, on the company’s WANDR evaluations, beat Opus 5 at one-fifth the price; Perplexity plans to use it as the “Light” effort orchestrator while Astra handles high-effort work.
The rollout therefore has two tracks: broadening access to cheaper AI agents, and a lingering challenge to show that the safety and alignment claims travel as reliably as the models do.
Write a comment