[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-88944bec-d33f-4383-aece-0d5207a06eab":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"88944bec-d33f-4383-aece-0d5207a06eab","GPT-5.6 全面开放:Ultra 把 4 agent 并行写进 API,程序化工具调用把 token 效率再压一档","7 月 9 日,OpenAI 把 GPT-5.6 三个档位 Sol\u002FTerra\u002FLuna 一起推向 GA。比起半年前 preview,真正落地的,是两个把「能力-成本曲线」再压一档的工程选择。\n\nUltra 模式——4 agent 并行编排。传统 scaling 在 thinking time 上做加法;GPT-5.6 走另一条路:4 个 agent 各管一段,主 agent 汇总。BrowseComp 单 agent 90.4%,Ultra 跑到 92.2%;Terminal-Bench 2.1 从 88.8% 抬到 91.9%。OpenAI 把它写进 Responses API 的 multi-agent beta,开发者可直接调,不用自建调度。\n\nProgrammatic Tool Calling(PTC)——工具调用从「透传」变「可编程」。过去每个 tool response 都要塞回模型下一轮;PTC 让模型写一段小程序,内部过滤聚合,只把下一动作送回主循环。早期客户数字:PlayCo 场景构建 token 砍 63.5%、turn 砍 50%;Clio 法律文档 prompt token 砍 38%、质量持平;Rogo 金融研究 output token 砍 24%、完成快 28%。\n\n更长上下文不是新闻(1M 早已 preview),但 GA 在 MRCR v2 8-needle 256K-512K 段跑到 91.5%,GraphWalks 1M BFS 77.1%——这一批旗舰里第一个把 1M 上下文从 benchmark 口号变成工程现实。配合 PTC,长上下文 agent 的成本模型第一次有了支撑点。\n\n拉远看,Sol\u002FTerra\u002FLuna 被定成「durable capability tiers」,会按各自节奏迭代,不再绑定某一代 release。配合三档定价,OpenAI 不再卖「最强模型」,而是卖「按预算可调度的算力梯度」——把「能力」切成可组合的运行时选项。\n\n对搭 agent 平台的团队,直接信号是:多 agent 编排从实验特性变 API 一档;工具调用从透传变可编程。这两个能力,第一次有了头部厂商的稳定接口。","https:\u002F\u002Fopenai.com\u002Findex\u002Fgpt-5-6\u002F","15975962-b5fe-49e5-ae68-687ba6cb7015",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"42e59a88-7795-47dc-a334-ef1e72c24347","openai","2026-07-10T06:03:00Z","2026-07-10T06:16:49.796795Z","2026-07-10T06:16:49.796808Z",true,"agent",2]