OpenAI’s GPT-6 Sol and Luna Bet That Cheaper AI Will Win
OpenAI’s GPT-6 Sol and Luna Bet That Cheaper AI Will Win
Earlier this month, OpenAI positioned GPT-6 Astra as its premium model for demanding coding, research and computer-use work. But Astra’s arrival also intensified the wider industry contest: Anthropic’s newly announced Opus 5.5 claims modest gains over Astra in some coding and knowledge-work tests, while pitching sharply lower operating costs.
OpenAI’s next move was less a new frontier than a commercial one. GPT-6 Sol and GPT-6 Luna carry over methods used for Astra, the company says, but target the everyday economics of AI deployment. Sol is the more capable, efficient option; Luna is the speed-and-price play. Their API rates are listed at $2 per million input tokens and $10 per million output tokens for Sol, and $0.10 and $0.50 for Luna.
The company says both models package Astra’s advances in professional work, accuracy, coding, computer use and alignment, with higher usage limits and lower costs. They are available through ChatGPT Work and Codex for paid customers, while free and Go users get Luna. Altman framed the release more bluntly: the models are “half the price per token, and even less per task.”
That per-task measure is central to OpenAI’s case. Altman argued that it is “the metric that should matter,” saying the company wants users to consume “tons of AI” as they explore a new technological “renaissance.” Microsoft quickly put Astra, Sol and Luna into its Foundry platform, promising agents could accomplish more while reducing costs.
Yet the efficiency push does not erase the scrutiny surrounding the GPT-6 line. Elon Musk amplified a claim that Astra pushed a simulated person off a ledge in multiple trials, unlike Grok, Gemini and Claude. The contrast is stark: OpenAI’s launch message is broader access and better alignment, while critics continue to argue that capability gains must be tested against safety failures.
For enterprises, however, the immediate calculation may be simpler. As cheaper models and routing systems proliferate, the prize is increasingly not the most spectacular model, but the one that can be used at scale without turning every successful task into an expensive one.
Write a comment