🔶 DeepSeek V4 正在赢得 Agentic 令牌份额
英文标题：🔶 DeepSeek V4 Is Earning Agentic Token Share
DeepSeek于4月24日发布旗舰V4模型，六个月内其在OpenRouter上的令牌份额接近翻倍，从年初的不到10%升至6月初的近20%，并一度成为最受欢迎的模型。增长主要由Agentic（代理型）工作负载驱动，V4 Flash的Agentic令牌量在一个月内占到DeepSeek代理型流量的70%。V4 Flash极具性价比（输入$0.09/百万令牌，输出$0.18/百万令牌），使得各类组织开始信任DeepSeek执行实际代理任务。这一趋势也推动中国开源模型整体令牌份额在2026年超越美国模型，对开发者意味着低成本高性能的Agentic模型正成为主流。
658.8 KB
456.3 KB
580.0 KB
492.6 KB
506.4 KB
488.3 KB
530.4 KB
429.6 KB
中文
DeepSeek，对于许多人来说依然是开源 LLM 的代名词，于 4 月 24 日发布了其新的旗舰 V4 模型。
V4 重新设定了发展轨迹。在六个月内，DeepSeek 在 OpenRouter 上的令牌份额大约翻了一番，而 Agentic 工作负载驱动了其中大部分的增长。
驱动当今令牌用量的模型作者
DeepSeek 年初在 OpenRouter 每周令牌流中的份额还不到 10%。随着 Agentic 工作在 2 月和 3 月开始占据主导，以前所未有的速度推高令牌使用量，DeepSeek 实际上跌至总令牌量的 5%。该公司受到了来自上方的闭源模型和来自下方的其他开源 LLM 浪潮的挤压。
V4 模型家族的发布立即产生了积极影响。到 6 月初，DeepSeek 已经赢得了近 20% 的令牌份额，并自 5 月中旬以来一直是 OpenRouter 上最受欢迎的模型。
2026 年 1 月与 6 月的直接对比显示了模型作者之间的偏好变化有多快。DeepSeek 在这一时期的令牌份额实际上翻了一番（从 9% 升到 18%），但许多其他模型提供商的命运也经历了重大波动。
一组中国开源模型，包括小米、MiniMax 和腾讯，在过去 6 个月中其令牌份额均有所上升。这似乎是以几家领先的美国模型公司为代价的，尤其是谷歌和 OpenAI。
DeepSeek 令牌用量的翻倍并不仅限于单一用户类型。到 6 月初，各类公司和个人使用的 DeepSeek 令牌量都在不断增加。业余爱好者用户（通常通过角色扮演、通用聊天、个人代理等面向消费者的应用类别产生大量使用量）现在将几乎三分之一的令牌路由到 DeepSeek 模型。
即使是人工智能原生公司或大型组织中的用户——他们本可能更倾向于前沿闭源模型——在 6 月初向 DeepSeek 发送的令牌流量也远高于年初。
DeepSeek V4 Flash，在最便宜的端点上，每百万令牌成本为输入 0.09 美元/输出 0.18 美元。相比之下，GPT-5.5 目前的定价为每百万令牌输入 5 美元/输出 30 美元。
因此，V4 令牌量的突然飙升并未导致支出份额出现同样幅度的增长，这也就不足为奇了。V4 的成本效益与输出质量之比是同类产品中最佳的——事实上，它足够优秀，以至于各种规模的组织都开始信任 DeepSeek 来承担真正的 Agentic 工作。
V4 是首个足以承担 Agentic 工作负载的 DeepSeek 模型
我们按照 API 密钥级别将令牌流量分为三个主要类别（Agentic、混合和人工）。这些类别是根据一个由 7 个信号组成的加权综合评分进行分配的，其中包括工具调用率、轮次数、间隔时间等。
众所周知，Agentic 工作比普通的人类 AI 使用消耗更多的令牌（根据 OpenRouter 的数据，每次请求大约多 15 倍）。但值得记住的是，这种 Agentic 令牌量的爆炸式增长是 2026 年的现象，实际上直到今年 2 月初才真正开始发力。
尽管对于某些模型作者来说，这种说法甚至还算保守。
Agentic 工作负载消耗的令牌量大约在 2 月 1 日左右超过了人工使用的令牌量。起初，Agentic 工作仅集中在少数几个模型作者身上，而 DeepSeek 并不在这个精选群体中。
V4 的发布标志着 DeepSeek 正式加入了 Agentic 工作流竞赛。
并非所有工作负载都会同时切换到最新模型。上图展示了在 DeepSeek 令牌使用量中，人类活动在整个年度前 5 个月仍与 DeepSeek V3.2 紧密相连。
然而，Agentic 工作负载却选择了 V4 作为他们的首选模型。到 5 月底，距离发布仅一个月，V4-Flash 就占据了 DeepSeek 使用中 Agentic 令牌流量的 70%。
DeepSeek 与中国模型的崛起
令牌流向两个国家创建的 LLM：美国和中国（少量其他国家的以浅褐色切片代表）。2026 年令牌量爆发式增长，但增长并未扩散到其他国家的 LLM。
2025 年是美国令牌之年，美国模型占据了使用令牌的大约四分之三。2026 年的竞争要激烈得多，到 6 月初，中国模型在令牌份额上实际已经超越了美国模型。
毫不意外，自 4 月下旬 V4 发布以来，DeepSeek 一直是中国模型中的领头羊。在过去 6 周，一些美国领先模型的令牌使用量趋于平稳，而随着 6 月的到来，多个中国竞争对手的令牌量出现了激增。
随着高性价比令牌使用成为夏季的头条新闻，我们预计这些成本更低但依然强大的开源模型将随着时间的推移继续吸引更多的令牌。正如《华尔街日报》在这篇近期报道（引用了 OpenRouter 的数据）中提到的，“初创公司及科技巨头都在混合搭配 AI 模型，以避免支付高昂的价格”。
DeepSeek V4 肯定会是这种混合组合的一部分。
方法论注释
关于我们如何得出这些数字的几点说明：
- 来源：OpenRouter 的请求日志。
- 样本量：2026 年 1 月 1 日至 2026 年 6 月 14 日期间超过 450 万亿个令牌。
- 份额指标：所有“份额”数据均指令牌量份额（输入与输出合计），而非支出份额。正如定价部分所指出的，对于像 V4 这样的低成本模型，两者可能相差甚远。
- 活动类型：我们将 API 密钥级别的令牌流量分为三个主要类别（Agentic、混合和人工）。这些类别是根据一个由 7 个信号组成的加权综合评分进行分配的，其中包括工具调用率、轮次数、间隔时间等。由于分类是在密钥级别进行的，执行混合工作的单个密钥将根据其主要模式进行分类，而非按每次请求进行。
English / 原文
DeepSeek, the company that for many is still synonymous with open source LLMs, released its new flagship V4 models on April 24th.
V4 reset the trajectory. In six months DeepSeek roughly doubled its share of tokens on OpenRouter, and agentic workloads drove most of that gain.
Model authors driving token usage today
DeepSeek began the year holding just under 10% of the token flow weekly across OpenRouter. As agentic work began to take hold in February and March, driving up token usage at unprecedented rates, DeepSeek actually fell to 5% of total tokens. The company was being squeezed from above by proprietary models and below from a wave of other open source LLMs.
The release of the V4 model family had an immediate positive impact. By the start of June, DeepSeek had earned nearly 20% of token share and has been the top model on OpenRouter since mid-May.
A direct comparison between January and June 2026 shows just how quickly preferences can shift between model authors. DeepSeek effectively doubled its token share over the period (from 9% to 18%) but many other model providers experienced major swings in fortune.
A group of Chinese open source models, including Xiaomi, Minimax, and Tencent, all saw their share of tokens rise over the past 6 months. This seems to have come at the expense of a couple of the leading American model companies, specifically Google and OpenAI.
The DeepSeek doubling was not restricted to a single user type. Companies and individuals of all stripes were using more and more DeepSeek tokens by the start of June. Hobbyist users (who typically have significant usage through consumer-oriented app categories like roleplay, general chat, personal agents, etc) now route nearly a third of all their tokens to DeepSeek models.
Even amongst users at AI native companies or large organizations, both of whom may have been expected to prioritize frontier closed models, sent DeepSeek much more token traffic in early June than they did to start the year.
DeepSeek V4 Flash, on the cheapest endpoint, costs $0.09 input / $0.18 output per million tokens. For comparison, GPT-5.5 is currently priced at $5 input / $30 output per million tokens.
So it’s not a shock to see this sudden upsurge in token volume for V4 has not resulted in an identical spike in share of spend. The cost effectiveness to output quality ratio for V4 is best in class - it is good enough, in fact, that organizations of all sizes have begun trusting DeepSeek with real agentic work.
V4 is the first DeepSeek model sufficient for agentic workloads
We segment token traffic at the API key level into three main categories (Agentic, Mixed, and Human). These are assigned according to a 7-signal weighted composite score that includes inputs such as tool call rate, turn count, gap timing, and others.
Everyone knows that agentic work burns more tokens than normal human AI usage (about 15x more per request, according to OpenRouter data). But it bears remembering that this explosion in agentic tokens is a 2026 phenomenon and really didn’t kick off in earnest until early February of this year.
Except even that is an overstatement for some model authors.
The tokens used by agentic workloads surpassed those used by humans right around February 1st. At first, agentic work was concentrated in only a few model authors and that select group did not include DeepSeek.
The release of V4 marked DeepSeek’s entrance into the agentic workflow race in earnest.
Not all workloads switch to the latest model at the same time. The chart above illustrates how human activity within DeepSeek token usage has remained tied to DeepSeek V3.2 through the first 5 months of the year.
Agentic workloads, however, have made V4 their model of choice. By the end of May, a mere month after release, V4-Flash comprised 70% of agentic token flow for DeepSeek usage.
DeepSeek and the rise of Chinese models
Tokens flow to the LLMs created in two countries: America and China (with a tiny smattering of others represented in the small beige slice). Token volume has exploded in 2026 but the growth has not spread to LLMs from other nations.
2025 was the year of American tokens, with models from the US responsible for about 3/4ths of the tokens used. The competition has been far more fierce in 2026, with Chinese models actually surpassing American ones in token share as of early June.
And to no surprise, it is DeepSeek leading the charge among Chinese models since the release of V4 in late April. While token usage amongst a few of the American leaders has plateaued over the past 6 weeks, multiple Chinese competitors saw tokens surge as June began.
With cost effective token usage dominating the summer headlines, we’d expect these cheaper-but-still-powerful open source models to continue to attract more tokens over time. As the Wall Street Journal mentions in this recent story (featuring OpenRouter data), “Startups and tech giants alike are mixing and matching AI models to avoid the premium prices”.
DeepSeek V4 is certain to be part of that mix.
Methodology Notes
A few notes on how we put these numbers together:
- Source : OpenRouter’s request logs.
- Sample size : Over 450 trillion tokens from January 1, 2026 - June 14, 2026.
- Share metric : All “share” figures refer to share of token volume (input and output combined), not share of spend. As the pricing section notes, the two can diverge sharply for low-cost models like V4.
- Activity type : We split token traffic at the API key level into three main categories (Agentic, Mixed, and Human). These are assigned according to a 7-signal weighted composite score that includes inputs such as tool call rate, turn count, gap timing, and others. Because the split happens at the key level, a single key doing mixed work is classified by its dominant pattern rather than per request.