OpenAI 推出 GPT-6 Sol 和 Luna 以及新的 API 定价

MarktechPost reports that OpenAI has released two new models in its GPT-6 family: GPT-6 Sol and GPT-6 Luna. These models sit below the GPT-6 Astra, which was launched earlier this month. OpenAI trained both models using methods similar to Astra, with the goal of making Astra's advances available in faster, more cost-effective models. Both models are already available in the OpenAI API under the names gpt-6-sol and gpt-6-luna. They are API-only models and therefore cannot be self-hosted. The GPT-6 family now consists of three levels. Astra is the top model for the most demanding tasks. Sol is aimed at complex coding and professional tasks at a lower price, while Luna is designed to quickly handle high-volume, everyday tasks. The OpenAI team reports that improved caching and inference make these models cheaper to serve. API prices for Sol and Luna are reduced by 50% compared to promotional prices for GPT-5.6. For example, the price for Sol is $2.00 per 1 million tokens, and for Luna it is $0.10. It's worth noting that the withdrawal price for Luna is reduced from $1.20 to $0.50, which is about a 58% reduction, not 50%. In professional performance tests

On AutomationBench 1.0.6, Sol scores 33.2% at best effort at $0.27 per task, which is better than Claude Opus 5's 26.9% at 11.1x the cost. On DeepSWE v1.1, Sol scores 68.8% at maximum effort, which is 1.1 points less than Claude Fable 5 at maximum effort, but the cost of the task is about 80% less. Luna at maximum effort scores 66.6%, which is comparable to Opus 5 and Fable 5 at medium effort, with Luna costing 93% less per task than Opus 5. OpenAI's internal fact-checking shows that Sol makes half as many errors as its predecessor. The Luna at higher forces matches the GPT-5.6 Sol at about 1/100th of its price. OpenAI has also carried over Astra's communication style, expecting clearer and slightly shorter responses with less jargon, especially in coding conversations. The request caching system has also been improved, allowing users to submit the same instructions, tools, and history at every step. The new Query Caching dashboard tracks hit rates over time and provides diagnostic tools. Both models are available

in API, ChatGPT Work and Codex, but not yet in Chat. The launch of ChatGPT will be gradual. The material was prepared by a neural network based on the original publication. AI may make mistakes in details (numbers, names, dates, quotes) - for important information, check with the original source. SpeakON launches magnetic voice input button Nokia opens up AnyJev: library for hands-free decision making Kyutai has released a Voice of Reason model for verbally solving math problems. Nokia opens access to AnyJev, a library for decision-making without training. Cache-to-Cache technique improves the interaction of language models, reducing latency.