GPT-5.6 Luna Reportedly Cuts API Prices by 80% as Agent Cost Competition Intensifies

Multiple reports say OpenAI has reduced GPT-5.6 Luna API prices by about 80% and Terra prices by 20%, while leaving Sol unchanged. DeepSeek has also opened public testing for V4-Flash-0731, shifting attention to model cost, performance, and agent capabilities.

Published 8/6/2026, 9:19:44 AMLast updated 8/6/2026, 9:19:44 AMEditorial publication

OpenAI has reportedly revised pricing across its GPT-5.6 API lineup. Luna’s input and output token rates have fallen to roughly one-fifth of their previous levels, Terra is about 20% cheaper, and Sol remains at the same price. Some reports also mention a faster API mode, although its exact terms should be confirmed in OpenAI’s documentation.

Around the same time, DeepSeek-V4-Flash-0731 was reported to enter public API testing. The available material describes improvements in selected agent benchmarks, along with support for the Responses API and Codex-related workflows.

The developments point to a broader competition around inference efficiency, pricing, and developer tooling. Lower model costs may improve the economics of agent products, but they do not by themselves guarantee profitability; workload complexity, infrastructure, caching, and operating expenses remain important variables.

Source evidence 6

Source review does not guarantee permanent accuracy. Consider the original source, observation time, and later corrections.

2026-08-01 AI 新闻汇总 - 智能体开发者社区adg.csdn.net · supporting

8 月 1 日，独立评测机构 Artificial Analysis 同步更新结果：V4-Flash-0731 智能指数 50 50 50 分，比 4 月原版高出 10 10 10 分，比 V4-Pro-Preview 高出 6 6 6 分，距 GPT-5.6 Luna 的 51 51 51 分仅差 1 1 1 分。在性价比方面，即使 OpenAI 当日将 Luna 价格下调 80 % 80\% 80%，V4-Flash-0731 在 DeepSeek 自有 API 上的单任务成本仍比 Luna 低约 60 % 60\% 60%，关键原因是 DeepSeek 提供约 98 % 98\% 98% 的缓存命中折扣，远超业内多数厂商的 90 % 90\% 90% 水平。Arena.ai 旗下 Frontend Code Arena（前端代码竞技场）显示，V4-Flash-High 以 1586 1586 1586 分排名前端代码领域第 7 7 7、开放类别第 3 3 3，每百万 token 价格 0.14 0.14 0.14 美元输入/ 0.28 0.28 0.28 美元输出，性价比居同类产品首位。 DeepSeek 同时透露，V4-Pro 正式版将尽快发布。该更新的产业含义在于：当模型架构趋于收敛，后训练与智能体执行框架的优化正成为性能提升的主要杠杆。一个不到 Pro 三分之一激活参数的 Flash 版本，仅靠后训练就在 Agent 能力上反超自家 Pro 预览版，说明数据质量与训练方法论的权重正在上升，单纯堆参数的路线开始被验证为非唯一解。 [...] # logo 智能体开发者社区 logo ## 登录社区云登录社区云，与社区用户共同成长 ### 智能体开发者社区邀请您加入社区 # 2026-08-01 AI 新闻汇总 ### nuo534202 ## 1. DeepSeek-V4-Flash 正式版 API 上线公测：仅凭后训练实现 Agent 能力反超 Pro 预览版，单任务成本比 GPT-5.6 Luna 低约 60 % 60\% 60% 7 月 31 日，DeepSeek 通过 API 文档发布日志宣布 DeepSeek-V4-Flash 正式版 API（版本号 V4-Flash-0731）进入公开测试。新版