OpenAI has introduced GPT-6 Sol and GPT-6 Luna, two additions to its GPT-6 family that the company positions as faster, lower-cost options for large-scale work. The models build on advances behind GPT-6 Astra, including what OpenAI describes as more efficient caching and inference.
Lower API prices
OpenAI claims GPT-6 Sol and Luna have API prices 50% lower than GPT-5.6 promotional pricing. The company also promotes higher usage limits, suggesting the models are intended to support more iteration at reduced cost.
Both models are available through the API. GPT-6 Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can try GPT-6 Luna in the desktop app.
Alignment claims
OpenAI also connects the new models to GPT-6 Astra’s alignment work, claiming that Sol and Luna improve on their GPT-5.6 counterparts.
The accompanying comparison materials list coding-deception rates of 1.3% for GPT-6 Sol and 2.8% for GPT-6 Luna, compared with 10.4% for GPT-5.6 Sol and 9.5% for GPT-5.6 Luna. GPT-6 Astra is listed at 0.5%. The chart labels lower rates as better, though the figures represent OpenAI’s stated evaluation results rather than an independent assessment.
Another AutomationBench comparison places GPT-6 Astra, GPT-6 Sol, GPT-5.6 Sol, GPT-6 Luna, GPT-5.6 Luna, Claude Opus 5, and Claude Fable 5.1 with Opus 5 fallback on a chart measuring score and cost per task. The chart uses score values from 0% to 50% and costs ranging from $0.005 to $5 per task.
Sol in ChatGPT
One accompanying ChatGPT interface identifies a model as “GPT-6 Sol Medium.” In an example involving a car wash 50 meters away, the model responds that driving is necessary because the car must be present for washing.
The launch materials also include a meme contrasting “GPT 6 SOL” with an attempt to remain loyal to “OPUS 5.5,” reflecting the familiar model-choice discussion around the release. OpenAI’s announcement, however, centers on cost, throughput, availability, and its claimed alignment improvements rather than a single benchmark result.
Source: OpenAI on X
