OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%


AI free token/AI免费积分
感谢IT之家网友 我家小妞很拽、兰花是我、华南吴彦祖 的线索投递!

IT之家 9 月 23 日消息,OpenAI 今日发布 GPT-6 系列模型的最新成员 GPT-6 Sol 和 GPT-6 Luna。

OpenAI 官方表示,两款模型使用与 GPT-6 Astra 类似的方法开展训练,将 Astra 在专业工作、事实性、编码、计算机使用和对齐等维度最先进的表现带入更快、更经济的模型。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

GPT-6 Sol 和 Luna 模型的 API 售价相比 GPT-5.6 促销价降低 50%。不过 OpenAI 官方也表示,GPT-6 Astra 仍旧是其总体上最出色的模型,如果你想要最佳成效和无妥协的感受,还是建议挑选 GPT-6 Astra。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

表现维度,在 AutomationBench 的跨应用业务流程测试中,GPT-6 Sol 在 xhigh 强度下表现优于 Claude Opus 5,成本仅为 Opus 5 每任务成本的 9%。在 high 强度下,GPT-6 Luna 比前代提高了 5.4%,且每项任务成本降低了 58%。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

在评估智能体的 Last Exam 测试中,GPT-6 Sol 在 max effort 状态下得分为 56.4%,高于 Claude Opus 5 在评估中的最高分,每任务成本也降低 60%。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

在事实可靠性维度,GPT-6 Sol 和 Luna 模型也取得进展。在 OpenAI 基于去标识化的真实世界对话中,GPT-6 Sol 的错误率约为前代的一半,接近 Astra 级可靠性且成本更低。GPT-6 Luna 的事实可靠性也迎来大幅提高,同时成本更低。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

编程维度,GPT-6 Sol 和 Luna 均针对 AI Coding Agent 工作负载开展了优化。在 FrontierCode 1.1 Main 测试中,GPT-6 Sol 相比 GPT-5.6 Sol 有显著提高,并以更低成本达到与 Claude Fable 5.1 xhigh 相当的水平。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

在 DeepSWE v1.1 软件工程测试中,GPT-6 Sol 在 max effort 下取得 68.8% 的成绩,距离 Claude Fable 5 的最高成绩 69.9% 仅差 1.1 个百分点,而单任务成本低约 80%。GPT-6 Luna 在 max effort 下取得 66.6%,表现接近 Opus 5 和 Fable 5 的 medium effort,同时单任务成本分别低约 93% 和 96%。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

计算机操作能力维度,OpenAI 表示 GPT-6 Astra 仍然是其在计算机操作维度最强的模型,但 Sol 和 Luna 相比上一代模型可以以更低成本完成相关任务。

在 OSWorld 2.0 offline 测试中,GPT-6 Sol 在 xhigh effort 下取得 60.5% 的成绩,与 Claude Opus 5 medium effort 的 60.3% 接近,而单任务成本约低 80%。GPT-6 Luna max effort 则超过 GPT-5.6 Sol medium effort,同时成本约为后者的颇为之一。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

OpenAI 还将 GPT-6 Astra 改进后的沟通风格带到了 Sol 和 Luna。官方认为新模型在工艺和编程对话中会更清晰、使用更少的术语、更少奇怪的措辞、更少的低价值细节,以及缩短总体回答,同时不会失去实质内容。

除了售价降低,OpenAI 还协助基于 GPT-6 的开发者在应用重复利用的上下文部分节省更多费用。官方改进了 GPT-6 的提示缓存,默认给出更高的缓存命中率,协助智能体重用更多上下文,达成更快响应。

另外,GPT-6 Sol 和 Luna 在价值对齐维度也有提高,在 OpenAI 官方的对齐评估中,Sol 和 Luna 都相较于 GPT-5.6 版本有所改进,涵盖关于其编码工作的误导性声明发生率降低。

OpenAI 发布 GPT-6 Sol 和 Luna 模型,API 价格大降 50%

IT之家从 OpenAI 公告获悉,GPT-6 Sol 和 GPT-6 Luna 自今日起已在 ChatGPT Work 和 Codex 中向所有 Plus、Pro、Business、Enterprise 和 Edu 使用者开放。免费使用者和 Go 使用者可以在桌面应用中使用 GPT-6 Luna。这些模型当下还没有在 Chat 中给出。在 OpenAI API 中,它们以 gpt-6-sol 和 gpt-6-luna 的形式给出。

本文来源:https://www.ithome.com/1/006/001.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Thanks to IT Home users for their clue submissions!

On September 23, OpenAI unveiled the latest members of the GPT-6 series: GPT-6 Sol and GPT-6 Luna.

According to OpenAI, these two models were trained using a method similar to GPT-6 Astra, bringing the latter’s cutting-edge performance in areas like professional work, factual accuracy, coding, computer use, and alignment to faster, more cost-effective models.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

The API prices for GPT-6 Sol and Luna are 50% lower than the promotional rate for GPT-5.6. However, OpenAI emphasizes that GPT-6 Astra remains its overall best model, so for maximum performance and uncompromising results, GPT-6 Astra is still the recommended choice.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

In terms of performance, in the AutomationBench cross-application workflow test, GPT-6 Sol outperformed Claude Opus 5 at xhigh intensity while costing only 9% of Opus 5’s per-task expense. At high intensity, GPT-6 Luna improved by 5.4% over the previous generation and reduced per-task cost by 58%.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

In the Last Exam test for evaluating agents, GPT-6 Sol achieved a score of 56.4% at max effort, exceeding the highest score recorded for Claude Opus 5 during evaluation, and also cut per-task costs by 60%.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

Regarding factual reliability, GPT-6 Sol and Luna have made significant progress. In OpenAI’s analysis of de-identified real-world conversations, GPT-6 Sol’s error rate was about half that of the previous generation, approaching Astra-level reliability at a lower cost. GPT-6 Luna’s factual reliability also saw a substantial boost with reduced cost.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

In the programming domain, both GPT-6 Sol and Luna were optimized for AI Coding Agent workloads. In the FrontierCode 1.1 Main test, GPT-6 Sol showed marked improvement over GPT-5.6 Sol and matched the performance of Claude Fable 5.1 at xhigh for significantly less.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

In the DeepSWE v1.1 software engineering test, GPT-6 Sol achieved 68.8% at max effort, just 1.1 percentage points behind Claude Fable 5’s top score of 69.9%, while per-task cost was approximately 80% lower. GPT-6 Luna’s max effort score of 66.6% closely matched Opus 5 and Fable 5’s medium efforts, with significantly reduced costs (up to 96% lower).

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

For computer operation capabilities, OpenAI notes that GPT-6 Astra is still the strongest in this area, but Sol and Luna can perform related tasks at lower costs compared to their predecessors.

In the OSWorld 2.0 offline test, GPT-6 Sol achieved 60.5% at xhigh effort, nearly matching Claude Opus 5’s medium effort score of 60.3%, with about 80% lower per-task cost. GPT-6 Luna surpassed GPT-5.6 Sol’s medium effort at a fraction of the cost.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

OpenAI has also extended GPT-6 Astra’s improved communication style to Sol and Luna, aiming for clearer, less jargon-filled, and more concise responses in technical and programming dialogues.

Beyond price reductions, OpenAI is helping developers save more by enhancing prompt caching for GPT-6 models, resulting in higher default cache hit rates and faster response times through better context reuse.

Additionally, GPT-6 Sol and Luna show improved safety alignment; OpenAI’s internal evaluations indicate lower rates of misleading self-assessments about their own capabilities compared to GPT-5.6.

OpenAI releases GPT-6 Sol and Luna models, API prices slashed by 50%

As reported by IT之家, starting today, GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. These models are not yet available in regular Chat; in the OpenAI API, they are identified as gpt-6-sol and gpt-6-luna.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry