OpenAI Counters With Cheaper GPT-6 Sol and Luna, Cutting API Prices in Half

OpenAI's Sol and Luna are trained with methods similar to GPT-6 Astra but tuned for lower cost and higher usage limits, aimed squarely at business and developer workflows where token efficiency, retries, and latency dominate economics. Sol is positioned as a production reasoning model at $2 input / $10 output; Luna is the ultra-cheap tier at $0.10 / $0.50. Sam Altman framed them as 'big improvements on intelligence, alignment, work output, coding, computer use' over the 5.6-family predecessors, and 'half the price per token, and even less per task.'

The timing was pointed: the models dropped within minutes of Anthropic's Opus 5.5 announcement, and the two launches together signal that low cost plus large capacity is becoming the entry ticket for frontier models rather than a differentiator. Independent developer Simon Willison called Luna 'astonishingly cheap' and his favorite model for building product features.

Separately (and covered in its own item), OpenAI on September 21 also published proposals urging the US to lead global technical standards for frontier AI alignment and recursive self-improvement. On the model front, availability is broad from day one — ChatGPT, Codex, Amazon Bedrock, and Microsoft Foundry all list Sol and Luna alongside Astra, letting customers match intelligence to task. The competitive read is that OpenAI is defending its low-cost turf, particularly against Chinese labs like DeepSeek whose entire pitch has been cheap inference. Community engagement leaned toward Anthropic on the day, but the GPT-6 launch still topped Hacker News at over 1,700 points.