GPT-5 发布

在 GPT-5 之前，许多人已经学会一种手工：简单问题用快模型，难题再切到推理模型。省心的一天还没到，选择本身先变成日常操作。菜单上的名字越来越多，账单与延迟却越来越需要人自己估。

2025 年 8 月 7 日，OpenAI 发布 GPT-5。ChatGPT 侧以路由系统按请求在快速模型与更深推理模型之间选择；API 侧则提供 GPT-5、mini、nano 等独立档位，以及可配置的推理强度，而不是把 ChatGPT 那套路由原样搬进每个请求。同一家族名下，消费产品与开发接口的分工不同——写混二者，会误判开发者实际拿到的控制权。用户在聊天里感到“系统替我决定”；开发者在 API 里仍要显式选档、设推理预算，才能得到可预期的延迟与账单。

路由的诱惑很具体：少做一次模型选择。代价同样具体：何时多花计算、何时直接答，变成平台策略与隐藏启发式的混合物。产品叙事可以说“系统替你决定”；工程账本仍要问“谁决定、依据什么、失败时能否覆盖”。GPT-5 也强化代码、工具调用与专业任务等定位，这些是发布方能力主张，应与具体评测设置一起读。

更耐记的变化或许不在某一张分数表，而在产品单位：一次回答之外，完整任务的成本、时延与是否调用工具，开始被写进默认体验。选择没有消失，只是从用户菜单部分上移到了系统路由——并在 API 里留下仍可手工拧动的旋钮。统一名称之下，分档与路由同时存在：省事与可控，被拆给了不同入口。

路由把省事送给用户，也把可解释性部分收走。出了错，人会问：刚才到底走了快路径还是慢路径？账单异常时，开发者会希望 API 侧的显式档位，而不是完全黑箱的自动选择。GPT-5 的产品史，因此是一次“智能分配计算”的实验，也是一次关于透明与便利如何并存的谈判。

快速与推理被装进同一家族名，并不等于用户理解了计费单位。产品文案若只说“更聪明”，账单明细仍会拆成不同路径的 token 与时间。路由史写到最后，仍是一笔关于可见性的账：便利增加时，解释是否同步增加。

Before GPT-5, many people had learned a manual habit: fast models for simple questions, a reasoning model when the problem hardened. The day of less fuss had not arrived; the choice itself had become daily work. Menu names multiplied; bills and latency still needed human estimation.

On 7 August 2025, OpenAI released GPT-5. On ChatGPT, a routing system selected between a fast model and a deeper reasoning model according to the request. On the API, separate GPT-5, mini, and nano tiers and configurable reasoning effort shipped instead of cloning the ChatGPT router into every call. Under one family name, consumer product and developer interface divided labor differently—confusing the two misreads how much control developers actually received. In chat, users feel the system decide; on the API, developers still pick tiers and set effort for predictable latency and bills.

Routing’s appeal is concrete: one fewer model choice. Its cost is concrete too: when to spend more compute and when to answer outright becomes a mix of platform policy and hidden heuristics. Product copy can say the system decides; the engineering ledger still asks who decides, on what basis, and whether failures can be overridden. GPT-5 also positioned stronger code, tool use, and professional work—publisher claims to be read with evaluation settings.

The more durable shift may not be one score table but the product unit: beyond a single answer, task cost, latency, and whether tools fire enter the default experience. Choice does not vanish; part of it moves from the user menu into system routing—and leaves knobs still turnable on the API. Under one name, tiers and routing coexist: ease and control are split across entrances.

Routing gives users ease and takes away part of explainability. When something fails, people ask: did that turn take the fast path or the slow one? When bills look wrong, developers want explicit API tiers rather than a fully black-box auto choice. GPT-5’s product history is therefore an experiment in “intelligently allocating compute,” and a negotiation over how transparency and convenience can coexist.

Putting fast and reasoning under one family name does not mean users understand billing units. If product copy only says “smarter,” invoices still split tokens and time across paths. Routing history ends as a ledger of visibility: when convenience rises, does explanation rise with it?