$4输入，$20输出——每百万Token。
8月21日，OpenAI宣布将GPT-5.6 Sol的API价格下调20%以上：输入价格从$5降至$4（降20%），输出价格从$30降至$20（降33%）。这次降价限时3个月，至11月21日，适用于API、ChatGPT Work credits和Codex credits，但不涉及Pro/Plus/Business订阅。
这已经是不到一个月内的第二次降价。7月底，OpenAI刚下调过中端和低成本模型的价格——中端模型降20%，低成本模型降幅甚至高达80%。如果说上一次降价是"清理中低端战场"，那这一次就是把价格战打到了自己的旗舰产品头上。
为什么是现在？三重压力叠加
前沿模型降价本身不是新闻，但OpenAI主动降价，而且是一月两降，这个节奏很值得琢磨。背后至少有三重压力。
第一重是Anthropic的正面竞争。Claude Sonnet 5发布后凭借速度和性价比抢走了大量开发者，很多企业从GPT-4级别模型直接迁移到了Sonnet。Anthropic Q2营收首次超过OpenAI——这个消息对OpenAI的冲击可能比外界想象的要大。开发者用脚投票的时候，价格是最直接的武器。
第二重是中国模型的价格碾压。DeepSeek、豆包、通义千问这些中国大模型的API价格，普遍在每百万输入token $0.2-$1的区间，和GPT-5.6 Sol的$4比起来，差了4-20倍。虽然性能有差距，但对于大多数任务来说，这个性能差距远没有价格差距那么大。很多海外开发者已经在悄悄用中国模型做非核心任务——OpenAI不可能视而不见。
第三重是推理成本的快速下降。降价的底气来自成本的下降——MoE架构、推理优化、芯片效率提升，让前沿模型的单位token成本一直在快速下降。既然成本降了，那就用降价换市场份额。这和消费电子行业的逻辑是一样的：摩尔定律让成本下降，厂商就用降价来刺激需求、扩大市场。
但需要注意的是：这次降价是限时3个月的，不是永久降价。为什么有限期？可能是暑期促销、可能是应对竞争的临时手段、也可能是想看看降价后的需求弹性——如果降价带来的用量增长足以抵消单价下降，那可能就永久降下去；如果不行，11月再涨回来也说不准。
价格战的终局：免费的基础层，赚钱的增值层
大模型的价格走势，其实和历史上很多基础设施技术的路径是一样的——从稀缺到过剩，从高价到低价再到接近免费。
早期，大模型是稀缺资源，谁有模型谁就能收钱。2023年初，GPT-4的API价格是$0.06/1k输入token（相当于$60/百万），比现在贵了15倍。那时候用大模型做产品的公司，最大的成本就是API调用费。
现在，模型能力在提升，价格却在下降。这意味着基础模型能力正在快速商品化——就像云计算的算力、移动互联网的带宽一样，会越来越便宜，直到接近成本价。未来，"调用一次大模型花多少钱"可能会像"发一条短信花多少钱"一样，不再是创业者最关心的问题。
那OpenAI靠什么赚钱？答案是：增值服务和生态锁定。你看微软的打法就很清楚——Azure云服务、Copilot产品、企业级部署，这些才是利润大头。模型本身可以便宜甚至免费，但一旦你进入了我的生态、用了我的工具链、接了我的Agent框架，你就很难走了。模型是入口，生态才是护城河。
对开发者和创业公司来说，这是个好消息：底层模型的成本不再是创业的主要障碍。你不需要再为API费用发愁，省下来的钱可以投入到产品设计、用户体验、垂直场景深耕上。未来AI创业的核心竞争力，不再是"我能调用某个独家模型"，而是"我对这个场景的理解足够深、我的产品体验足够好"。
但反过来说，这也意味着仅靠套壳大模型的创业公司会死得更快。如果模型能力本身快速商品化，那你的价值是什么？如果只是把用户的query转发给大模型再把答案转回来，那这个中间层会被迅速挤压掉。要么你有独家数据，要么你有独特场景，要么你有产品壁垒——否则，在模型越来越便宜的时代，纯中介没有生存空间。
价格战才刚刚开始，后面还会有更多轮降价。对于整个行业来说，这是走向成熟的必经之路；对于每个玩家来说，这是考验真正价值的时候。
明天见。
$4 input, $20 output — per million tokens.
On August 21, OpenAI announced it was cutting GPT-5.6 Sol API pricing by over 20%: input dropped from $5 to $4 (20% off), output from $30 to $20 (33% off). The cut is limited to 3 months, through November 21, and applies to the API, ChatGPT Work credits, and Codex credits — but not to Pro, Plus, or Business subscriptions.
This is already the second price cut in less than a month. Back in late July, OpenAI lowered prices on mid-tier and low-cost models — 20% off mid-tier, and a staggering 80% off low-cost ones. If the last round was "cleaning up the mid-to-low end battlefield," this one is bringing the price war to its own flagship product.
Why Now? Three Pressures Stacking Up
Frontier model price cuts aren't news by themselves, but OpenAI cutting them proactively — and twice in one month — is a pace worth examining. There are at least three pressures behind it.
First, direct competition from Anthropic. Since Claude Sonnet 5 launched, it's stolen massive developer share on speed and price-performance. Many enterprises migrated directly from GPT-4-class models to Sonnet. Anthropic's Q2 revenue surpassing OpenAI for the first time — that news probably hit harder internally than anyone admits. When developers vote with their feet, price is the most direct weapon.
Second, price pressure from Chinese models. Chinese LLMs like DeepSeek, Doubao, and Tongyi Qianwen generally price in the $0.20–$1 per million input tokens range — a 4x to 20x difference from GPT-5.6 Sol at $4. While there's a performance gap, for most tasks the gap isn't nearly as big as the price gap. Plenty of overseas developers are quietly using Chinese models for non-critical tasks. OpenAI can't ignore that.
Third, rapidly falling inference costs. The confidence to cut prices comes from falling costs — MoE architectures, inference optimization, chip efficiency gains keep driving unit-token costs down for frontier models. Since costs are dropping, why not trade price for market share? It's the same logic as consumer electronics: Moore's Law drives costs down, manufacturers use price cuts to stimulate demand and expand the market.
But it's worth noting: this price cut is time-limited to 3 months, not permanent. Why the expiration? Maybe it's a summer promotion, maybe a temporary competitive move, maybe a test of demand elasticity — if increased usage from lower prices offsets the per-unit drop, it might become permanent. If not, prices could go back up in November. We'll see.
The Price War Endgame: Free Base Layer, Paid Premium Layer
The price trajectory of LLMs actually follows the same path as many infrastructure technologies in history — from scarcity to abundance, from expensive to cheap to nearly free.
In the early days, LLMs were a scarce resource — whoever had a model could charge for it. In early 2023, GPT-4 API pricing was $0.06/1k input tokens (equivalent to $60/million) — 15x more expensive than now. Back then, the biggest cost for companies building LLM products was API calls.
Now, model capabilities are rising while prices fall. This means base model capability is rapidly commoditizing — just like cloud computing power or mobile internet bandwidth, it'll get cheaper and cheaper until it approaches cost. In the future, "how much does an LLM call cost" might be as irrelevant to founders as "how much does an SMS cost" is today.
So how does OpenAI make money? The answer: premium services and ecosystem lock-in. Look at Microsoft's playbook — Azure cloud services, Copilot products, enterprise deployment — those are where the real profits are. The model itself can be cheap or even free, but once you're in my ecosystem, using my toolchain, plugged into my agent framework, you can't easily leave. The model is the entry point; the ecosystem is the moat.
For developers and startups, this is good news: base model cost is no longer the main barrier to entry. You don't need to stress about API bills anymore. The money you save can go into product design, user experience, deep vertical expertise. In the future, the core competitive advantage in AI startups won't be "I have access to some exclusive model" — it'll be "I understand this scenario deeply enough and my product experience is good enough."
But conversely, it also means pure wrapper startups will die faster. If model capability itself is rapidly commoditizing, what's your value? If you're just forwarding user queries to an LLM and forwarding answers back, that middle layer gets squeezed out quickly. Either you have proprietary data, a unique scenario, or product barriers — otherwise, in an era of cheaper and cheaper models, pure middlemen have no生存 space.
The price war has only just begun. There will be many more rounds of cuts. For the industry as a whole, this is the inevitable path to maturity. For each individual player, it's a test of real value.
See you tomorrow.
Base model capability is rapidly commoditizing — just like cloud compute or mobile bandwidth, it gets cheaper and cheaper until it approaches cost.
— The Dawn Vision Editorial Desk
GPT-5.6 Sol · OpenAI price cut · LLM price war · model commoditization · Anthropic competition · Chinese model pricing · falling inference costs
