文 | 影子备忘录
OpenRouter最新一周的数据显示。10月19日到25日，DeepSeek V4.1-Flash-1017以33.6万亿Token的周调用量登顶全球第一，环比增长801%。
中国模型包揽前五，周总调用量93.54万亿Token，连续26周压过美国的22.47万亿。
但如果你只看到“第一”两个字，大概率会误判这件事的真实分量。
33.6万亿Token是一个实打实的数字，但它不是“技术碾压”的结果，它仅仅只是一场“账单碾压”的阶段性兑现。
更准确地说，这是全球开发者和中小企业在真实工作流里，用每一次API调用做的成本决策。理解这一点，比庆祝排名本身重要得多。
谁更会算账？
V4.1-Flash-1017是10月17日才发布的版本。同一周，上一代V4.1 Flash的调用量从更高水位掉到18万亿Token，环比下滑45%。
老版本的用户向新版本迁移，新版本接住了几乎全部存量需求，再叠加少量真正的增量开发者。一进一出，环比数字被急剧放大。
版本迁移本身说明：用户对价格和性能的变化极其敏感，迁移成本足够低，切换决策几乎是即时的。
一个模型能在发布后一周内完成如此规模的新旧交替，恰恰证明它的用户基本盘是活跃的、在意的、随时准备“用脚投票”的。
但同样需要正视的是，OpenRouter并不是全球大模型调用量的全景图，它统计的是经由这个API聚合网关中转的流量。
厂商自有API、Azure和Bedrock上的企业级部署、私有化推理，以及直接面向消费者的App和网页产品，都不在这个池子里。
33.6万亿Token证明的是“在这个开放聚合平台上，DeepSeek是开发者最愿意调用的模型”，并非“全球AI市场已经被DeepSeek统治”。
把这两件事分开看，才能继续往下追问：为什么偏偏是它，在这个平台上被如此密集地调用？
答案不在参数表的最上面一行，而在开发者的成本核算表里。
V4.1-Flash的总参数是552B，但Prefill阶段每个Token只激活8B，Decode阶段激活16B。
这意味着它的“脑容量”接近旗舰级，但每次推理真正动用的计算资源被压到了极低水位。
支撑这个效率的是一套并不常见的架构选择：40层语言主干被拆成20层因果编码器和20层Decoder，Prompt在前半段完成编码后，后半段需要的全局KV直接从编码器末端投影得到，不需要让全部历史Token反复穿过整个网络。
官方报告的说法是，Prefill计算量接近减半。
更关键的是KV Cache的处理。Agent工作流的特点是“读得远远多于写”，每一次工具调用都会灌入新的网页、代码、终端输出，上下文迅速膨胀。
如果每次都要重新计算全部历史，成本会线性甚至指数级上升。
DeepSeek的做法则是把KV Cache从单机内存里抽出来，放进独立的3FS分布式存储集群，缓存容量约100 PiB，Agent流量保留3天，聊天流量保留7天。
重复的上下文不再反复烧算力，这是Agent场景下推理经济性的结构性改善。
这些技术细节最终指向同一个结论：在“够用”和“可负担”之间，DeepSeek把平衡点推到了一个让大量中小企业无法拒绝的位置。
有测算显示，Anthropic头部模型完成同类Agent任务的成本，大约是它的28倍。
V4.1-Flash-1017在性能提升的同时维持原价，即每百万输入Token 0.15美元、输出0.60美元，MIT协议开源。
更早一轮的定价调整中，DeepSeek全系列API的缓存命中价格被压到原价的十分之一，V4-Flash的缓存命中输入价格降至0.02元/百万Token。
这不是真正形式上的“便宜”，因为它是用架构效率换出来的定价空间。
当大量真实工作流，诸如客服分类、文档打标签、代码补全、Agent工具调用等，对“绝对智能上限”的敏感度远低于对“每次调用成本”的敏感度时，全球开发者用Token投票的方向几乎是注定的。
开源的飞轮，比降价本身更难被复制
如果只是便宜，DeepSeek的登顶可以被理解为一场价格战的短期结果。
但价格战谁都能打，真正让“便宜”变成结构性优势的，是开源把定价权从厂商手里部分转移到了用户手里。
10月20日，DeepSeek开源了内部推理框架DeepInfer，同时兼容英伟达和华为昇腾两大算力平台。
这件事的意义不在于“又多了一个开源项目”。在此之前，即使企业拿到了V3、R1的模型权重，也不等于能以官方同等效率跑起来。
推理引擎、并行策略、缓存管理、调度逻辑……这些才是把模型权重转化为低成本服务的“暗知识”。
DeepInfer把这些暗知识公开了：任何企业都可以在自己的集群上部署接近官方水准的推理服务，中小企业不必从零搭建优化能力，高校团队可以在昇腾上验证自己的设计。
这形成了一个别人很难用一次降价来打断的循环：开源权重降低了采用门槛，开源推理框架降低了部署门槛，部署规模扩大又反过来喂养模型迭代和成本优化。
当中国模型的全球下载量已经占到开源平台的大头，当V4的百万Token上下文和权重面向全球开发者免费开放，所谓“封堵”在软件层面的有效性就在持续衰减。
你可以禁止一个机构账户，但你很难禁止一套已经进入全球开发者工具箱的权重和框架。
不过话说回来，我们看待问题，还需要更全面才行。
毕竟OpenRouter上Token多，不等于利润多。
在同一个平台上，按任务花费金额排序，占据高价值分类任务头部位置的仍然是Claude Opus 5.5、GPT-5.6 Sol和Gemini 3.1 Pro，唯一进入分类任务花费前列的中国模型才是DeepSeek V4.1 Flash。
高复杂度、高风险、高价值的工作流，企业仍然愿意为“最聪明的那一个”支付溢价。DeepSeek拿下的，是真实工作流中体量最大、但单位价值相对较低的那一层。
这决定了它的商业处境并不像调用量排名那样从容。有分析指出，其实际混合毛利率远低于软件行业的正常水位，200人左右的团队规模和头部闭源实验室数千人、百亿美元级收入之间的差距，也不是一次登顶能够填平的。
API价格在部分阶段已经出现回调节奏，“极致低价换规模”正在让位于“规模之上怎么算单位经济模型”。
调用量是护城河的一部分，但护城河本身不产生利润。
技术原创性上的质疑同样没有消失。蒸馏的指控反复出现，尽管DeepSeek此前的R1论文曾以训练数据来源的可验证性做出过回应，但围绕底层架构“原创程度”的争论不会因为一次周榜第一而终结。
更现实的约束来自算力：向国产芯片迁移是必须走的一步，单卡性价比可以算账，但万卡规模的通信稳定性和下一代模型的预训练效率，才是决定它明年是否还在榜上的硬变量。
地缘层面的压力也在收紧。美国国土安全部已要求联邦文职机构从IT系统中移除DeepSeek产品及基于其模型开发的第三方服务，限制范围从国防和情报领域扩展到了近乎全部联邦民事机构。
开源可以绕开一部分商业封锁，但绕不开一个国家层面的制度性排斥。
结语
过去一年，OpenRouter上的中国模型份额从个位数一路推到与海外模型分庭抗礼，全球开发者的Token流向发生了方向性变化。
这个变化的驱动力不是某一项“碾压性”的技术突破，而是一条更朴素、也更难被短期复制的能力链：架构效率带来成本空间，成本空间支撑定价，定价吸引开发者，开源把开发者锁定为生态参与者，生态再反哺迭代。
但这条链的每一环都有它的边界。调用量的领先可以靠性价比赢得，利润的领先要靠不可替代性赢得，技术代际的领先要靠原创性赢得。
DeepSeek赢下了第一张卷子，不等于三张卷子都答完了。
33.6万亿Token是真实的，全球开发者的选择是理性的，但“全球第一”的数据之下，真正值得记住的是那个让排名发生的原因：
在一个越来越贵的行业里，有人选择把“用得起”做成核心竞争力。
这条路能走多远，取决于它能不能在“便宜”之外，继续给出“非它不可”的理由。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The Shadow Memorandum
OpenRouter's latest week data. From 19 to 25 October, DeepSeek V4.1-Flash-1017 came to the top of the world with 33.6 trillion Token weekly calls, an 801 per cent increase in the ring.
The Chinese model pack was first five years, with a total weekly call of 93,54 trillion Token and 22,47 trillion dollars over 26 consecutive weeks in the United States.
But if you only see the first word, you probably misjudge the true weight of the matter.
Token, 33.6 trillion, is a solid figure, but it is not the result of “technical crushing”, but only a “billing” phase.
More precisely, it is a cost decision made by global developers and SMEs in real workflow using each API call. Understanding this is much more important than celebrating the ranking itself.
Who's better at accounting?
V4.1-Flash-1017 was only publish on 17 October. In the same week, the previous generation, V4.1 Flash, fell from a higher water level to 18 trillion Token, a 45 per cent decline in the ring.
The old version of the user migrations to the new version, which captures almost all inventory requirements, with a few real incremental developers. Once in and out, the ring numbers are dramatically amplified.
The version itself indicates that users are extremely sensitive to changes in prices and performance, that the costs of migration are low enough and that switching decisions are almost instantaneous.
The fact that a model can complete a new and old rotation of this magnitude within one week after publish is proof that its user base disk is active, conscious and ready to vote “by foot”.
But it is also important to face up to the fact that OpenRouter is not a global LLM call map, and it counts the traffic through this API aggregation gateway.
The plant does not have an API, Azure and Bedrack enterprise-level deployment, privatization reasoning, and the App and web-based products directly aimed at consumers.
Token, 33.6 trillion, attests to the fact that “DeepSeek is the most preferred model for developers on this open polymer platform” and not “the global AI market has been ruled by DeepSeek”.
If we look at these two things separately, we can continue to ask: why is it so heavily called on this platform?
The answer is not in the top line of the parameter table, but in the developer ' s cost statement.
The overall parameters of V4.1-Flash are 552B, but only 8B, Decode 16B is activated per Token in Prefil.
This means that its “brain capacity” is close to the flagship, but the computational resources actually used for each reason are pressed to a very low level.
This efficiency is underpinned by an unusual set of structural choices: 40-storey language backbones are broken down into 20-storey causal encoders and 20-storey Decoder, and when Prompt completes the coding in the first half of the first half, the global KV required in the second half is projected directly from the end of the encoder, without the need for Token of all history to cross the entire network.
According to the official report, Prefil counts are close to halving.
More crucial is the handling of KV Cache. Agentworkflow is characterized as “reading far more than writing”, each time Z17Z is inserted into a new web page, code, terminal output and context is expanding rapidly.
If the entire history is recalculated each time, the cost is linear or even exponential.
DeepSeek's approach is to remove KV Cache from a single memory and place it in an independent 3FS distributed storage group, cache capacity approximately 100 PiB, Agent traffic remaining for 3 days and chat traffic remaining for 7 days.
The repetition of context is no longer repeated calculus, a structural improvement in the economics of reasoning in the Agent scene.
These technical details ultimately point to the same conclusion: between “sufficient” and “affordable”, DeepSeek pushed the balance point to a position where a large number of SMEs could not be rejected.
Measurements indicate that the cost of the Anthropic head model to perform the same type of Agent mission is approximately 28 times greater.
V4.1-Flash-1017 maintains the original price for performance enhancement, i.e. token 0.15 per million, output $0.60, MIT protocol open source.
In the earlier round of pricing adjustments, DeepSeek ' s cache hit price of the whole series of API was pressed to one tenth of the original price, and the cache hit input price of V4-Flash fell to 0.02 yuan/million Token.
This is not a real form of “cheap”, because it is a price-fixing space that has been replaced by structural efficiency.
When a large number of real workflows, such as customer classification, document labelling, code completion, Agenttool call, are much less sensitive to “absolute smart caps” than to “per call costs”, it is almost certain that the global developers will use Token to vote.
Open-source wheels are harder to copy than lower prices per se.
If it's cheap, DeepSeek's summit can be understood as a short-term result of a price war.
But what really transforms “cheap” into a structural advantage is that the open source shifts pricing from the manufacturer to the user.
On 20 October, DeepSeek opened the internal framework of reasoning, DeepInfer, and compatible with Britain and China ' s two major computing platforms.
The point is not “an additional open source project”. Prior to this, even if an enterprise had a model weight of V3, R1, it would not have been able to run with the same official efficiency.
Educate engines, parallel strategies, cache management, movement logic These are the “dark knowledge” that transforms model weights into low-cost services.
DeepInfer made these dark knowledges public: any enterprise could deploy a reasoning service close to the official level on its own clusters, SMEs did not have to build their optimization capacity from a patchwork, and higher education teams could test their design on a stagger.
This creates a cycle that is difficult for others to interrupt with a one-time price reduction: open source weights lower the threshold for adoption, open source reasoning frameworks lower the threshold for deployment, and larger deployments which in turn result in feeding models that are iterative and cost optimization.
The effectiveness of the so-called “blockage” at the software level continues to decline when the global download of the Chinese model is already at the top of the open source platform, when the million Tokencontext and weights of V4 are freely open to global developers.
You can ban an institutional account, but it's hard for you to ban a set of weights and frameworks that are already in the global developers' toolbox.
But then again, we need to look at the problem more comprehensively.
After all, OpenRouter has more tokens than profit.
On the same platform, in order of the amount spent on the task, Claude Opus 5.5, GPT-5.6 Sol and Gemini 3.1 Pro continued to occupy positions in the head of the high-value classification mission, and DeepSeek V4.1 Flash was the only Chinese model that entered the top of the cost of the classification mission.
High-complex, high-risk, high-value workflow, businesses remain willing to pay premiums for “the smartest”. DeepSeek took the largest but relatively low unit value of the real workflow.
This determines that its business position is not as acceptable as the call ranking. The analysis indicates that the actual MMR is well below the normal water level in the software industry, and that the gap between the size of the team of about 200 people and the tens of billions of dollars in head closed source laboratories is not a single peak that can be filled.
The API price has been re-regulating in part, and “extremely low-price-for-scale” is giving way to “how to calculate a unit economic model above scale”.
The mobilization was part of the moat, but the moat itself did not generate profits.
Technically original questions have likewise not disappeared. The allegations of distillation have been repeated, and although DeepSeek's previous R1 thesis had responded with the probabilities of training data sources, the debate over the “originality” of the bottom structure did not end with the first of a week.
A more realistic constraint comes from calculus: moving to a national chip is a necessary step, and single-card value can be calculated, but the stability of communications on the scale of 10,000 cards and the pre-trained efficiency of the next generation model is the hard variable that determines whether it will be on the list next year.
The pressure at the geologic level is also tightening. The United States Department of Homeland Security has requested the federal civil service to remove DeepSeek products from the IT system and third-party services based on its model, with restrictions extending from defence and intelligence to almost all federal civil agencies.
Open sources can bypass some of the commercial blockades, but not a country-level institutional exclusion.
Concluding remarks
Over the past year, the share of the Chinese model on OpenRouter has been pushed from single digits to resistance to overseas model chambers, and the global developer Token's flow has changed in direction.
The driving force of this change is not a technological breakthrough of a “crushing” nature, but a chain of capabilities that is more simple and more difficult to replicate in the short term: architecture efficiency brings with it cost space, cost space support pricing, pricing attracts developers, open source locks developers into ecological players, and ecological re-energizing.
But every ring of this chain has its borders. The lead in mobilizing amounts can be won by value for money, the lead in profits by non-alternatives, and the lead in technology generation by originality.
DeepSeek won the first one, not all three of them.
The 33.6 trillion Tokens are real, and the global developers' choices are rational, but under the “global first” data, it's really worth remembering what made the ranking happen:
In an increasingly expensive industry, “affordability” has been chosen as a core competitiveness.
How far this road can go depends on whether it can continue to give “necessarily” reasons beyond “cheap”.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.