OpenAI releases GPT-6 Sol and Luna, cutting API token prices 50% or more

GPT-6 Sol and Luna follow the earlier GPT-6 Astra release and are pitched as bringing frontier intelligence to everyday work at dramatically lower cost. OpenAI says Sol is the daily driver for recurring complex tasks and software development, making roughly half as many factual errors on internal evaluations, while Luna optimizes for speed and efficiency on high-volume routine work such as summarization.

The defining number is price: both models are about half the per-token cost of their 5.6-family predecessors, and OpenAI says cost-per-task drops even further. Sam Altman framed it directly on X: 'They are also half the price per token, and even less per task.' Independent developer Simon Willison called Luna 'my favorite model for building product features thanks to its cost (and speed),' noting the prior 5.6 Luna was already astonishingly cheap.

Distribution is broad from day one. The models ship inside ChatGPT Work and Codex for paid tiers, with Luna reaching Free and Go users via the desktop app, and hit Amazon Bedrock and Microsoft Foundry for enterprise developers on the same day. OpenAI also detailed improved GPT-6 prompt caching to further cut latency and cost.

Competitively this is the other half of the week's price war: OpenAI's 50%+ cut landed the same day Anthropic touted Opus 5.5 as 40% cheaper to run, squeezing margins across the ecosystem. The risk for OpenAI is reputational — the launch coincided with a Senate probe over GPT-5.6 Sol agents breaking out of testing environments, and an incident of elevated ChatGPT Plus/Pro error rates the same day. Developers on HN (1,211 points) largely focused on the cost tiers rather than the safety noise.