OpenAI于7月31日宣布GPT-5.6系列模型大幅降价[1]。其中Luna模型降幅最大，输入价格从1美元下降至0.2美元，输出价格从6美元降至1.2美元，总体降幅达80%[1]。这一定价已经低于DeepSeek V4 Pro的0.435美元/百万Token[1]。Terra模型的输入价格也从2.5美元降至2美元，输出从15美元降至12美元，降幅为20%[1]。相比之下，Sol模型价格保持不变，输入和输出分别维持在5美元和30美元，但新增Fast模式使其速度提升2.5倍[1]。
这次大幅降价得益于通过优化基础设施实现的成本下降。OpenAI重写了GPU内核、改进推测解码使其效率提升15%、优化了负载均衡并调优了KV缓存配置，从而使服务成本下降20%以上[1]。OpenAI首席执行官Sam Altman表示要在"每一个模型层级都提供最好的价格与智能比"[1]。同日，DeepSeek发布新版本，声称代码编写和调试能力大幅提升[1]。根据OpenRouter数据，自2026年4月下旬以来，中国模型Token调用量已连续十三周超过美国模型，7月20-26日期间中国模型处理了38.6万亿Token，占全平台66.5%[1]。
OpenAI announced dramatic price reductions for its GPT-5.6 series models on July 31, with the Luna model's input pricing cut by 80% from $1 to $0.2 per million tokens, positioning it below DeepSeek's offerings [1]. The Luna model's output price also fell from $6 to $1.2 per million tokens, while the Terra model saw input costs drop from $2.5 to $2 and output from $15 to $12, representing a 20% reduction [1]. The Sol model maintained its pricing at $5 for input and $30 for output but introduced a new Fast mode delivering 2.5 times faster processing speeds [1].
The price cuts were enabled by infrastructure optimizations in the Sol model through Codex, which included rewritten GPU kernels, improved speculative decoding with 15% efficiency gains, optimized load balancing, and refined KV cache configurations, collectively reducing service costs by over 20% [1]. OpenAI CEO Sam Altman stated the company aimed to provide "the best price and intelligence ratio at every model tier" [1]. On the same day, DeepSeek released a new version claiming substantial improvements in code writing and debugging capabilities [1].
Luna's input pricing of $0.2 per million tokens now undercuts DeepSeek V4 Pro at $0.435 per million tokens [1]. According to OpenRouter data, Chinese AI models have dominated token processing for thirteen consecutive weeks since late April 2026, handling 386 trillion tokens in the week of July 20–26, representing 66.5% of the platform's total volume [1].
