OpenAI在GPT-6 Sol发布仅7天后推出了新一代模型GPT-6.1 Sol1。新模型在智能指数上相比前代提升4分，仅低于旗舰产品GPT-6 Astra 1分，但成本不到后者的四分之一1。
GPT-6.1 Sol在多项基准测试中表现显著提升，其中AA-Omniscience基准在最大努力设置下跳升8分，幻觉率从60%降至54%1；Terminal-Bench 4.0提升12分，Humanity's Last Exam提升5分，GDP.pdf提升6分1。在编码能力方面，编码代理索引相比GPT-6 Sol提升3分，但低于GPT-6 Astra 2分1。
就成本而言，GPT-6.1 Sol的单任务成本为$0.72，相比GPT-6 Sol的$1.05便宜31%，相比GPT-5.6 Sol的$1.99便宜64%1。模型价格维持每百万令牌$2/$10（输入/输出），缓存读取折扣从90%提升至95%1。
OpenAI has released GPT-6.1 Sol, a new model that succeeds GPT-6 Sol merely seven days after its debut 1. The upgraded version achieves near-parity with GPT-6 Astra in capability while maintaining substantially lower costs, positioning itself as a significant middle ground in the company's model lineup.
The new model delivers notable performance improvements across multiple benchmarks. GPT-6.1 Sol surpasses its predecessor by four intelligence index points and trails GPT-6 Astra by only one point 1. On specialized evaluation metrics, GPT-6.1 Sol jumps eight points higher than GPT-6 Sol on AA-Omniscience under maximum effort settings, while reducing hallucination rates from 60 to 54 percent 1. The model also achieves substantial gains on Terminal-Bench 4.0, with a 12-point increase, and demonstrates progress on Humanity's Last Exam and GDP.pdf, improving by 5 and 6 points respectively 1.
From a cost perspective, GPT-6.1 Sol represents a compelling value proposition. Under maximum effort settings, the per-task cost stands at $0.72, roughly one-quarter the expense of GPT-6 Astra at $3.26 and 31 percent cheaper than the previous GPT-6 Sol model at $1.05 1. OpenAI has maintained the input and output token pricing at $2 and $10 per million tokens respectively, while raising the cache read discount from 90 to 95 percent 1. In coding agent performance, the model lags GPT-6 Astra by two points under maximum effort, though it still shows a three-point improvement over GPT-6 Sol 1.
评论
还没有评论，欢迎留下第一条。