闭源Token价格指数跌至3.07美元、开源仅0.66美元,token交易市场 中美算力成本差距加速收窄

闭源集体降价让利、开源价值稳步抬升,全球 Token 定价体系正在重塑。


The $3.07 Closed-Source Price vs. the $0.66 Open-Source Price: Global Token Pricing Is Converging

English Version

In early August 2026, a set of industry data released by Silicon Data caused a stir across the AI sector. The comprehensive token price index for closed-source models dropped to $3.07 per million tokens**, while the index for open-source models rose to **$0.66 per million tokens.

Two pricing curves are advancing toward each other’s original track: closed-source model prices are trending downward from $3.07, while open-source model prices are steadily climbing from $0.66. The current price gap between them has narrowed to approximately 4.65 times.

This is no longer the old narrative of “open source is cheaper than closed source”. It represents a brand-new industry pattern: closed-source prices are falling, and open-source prices are rising.

Two Core Forces Behind the Pricing Shift

First, it is necessary to clarify the definition of the Silicon Data LLM Token Expenditure Index. It calculates the weighted average of the actual market expenditure per million tokens across mainstream models, reflecting real user spending rather than official nominal listing prices.

Closed-source prices are declining sharply driven by fierce market competition. OpenAI slashed the pricing of its Luna model by 80%; Anthropic canceled a planned 50% price hike for Sonnet 5; Google cut the price of Gemini 3.7 Flash by half. The three global tech giants launched large-scale price cuts simultaneously. In the current homogeneous competition landscape, any overpriced API service will be quickly eliminated, as enterprise CIOs will immediately optimize costly model solutions.

In contrast, open-source token prices are steadily rising. Although open-source model weights are free of charge, the actual inference operation incurs substantial costs, including GPU computing power, engineer parameter tuning, quantitative optimization, and 24/7 operational maintenance and troubleshooting. Free weights do not equal free deployment.

More importantly, the market share of open-source models has exploded. Chinese open-source models’ occupancy on OpenRouter surged from less than 2% in early 2025 to 46% in June 2026. The sharp growth in real usage has directly pushed up the comprehensive operational cost of open-source services.

The two curves are gradually converging in the mid-price range. Closed-source price cuts stem from market competition, while open-source price increases come from explosive user demand. This dual trend is rapidly narrowing the historical computing power cost gap between China and the United States.

Underlying Logic of Narrowing Sino-US Computing Power Cost Gap

The shrinking price gap is not merely a numerical change, but a fundamental overturn of global AI pricing logic.

In the past, Chinese model manufacturers dominated the market relying on extreme cost performance. For example, DeepSeek V4 Pro charges only $3.96 per million output tokens during peak hours and offers a 50% discount for off-peak periods, while the overseas Fable 5 is priced at $50 per million tokens — a gap of more than 10 times.

This huge gap is now being rapidly compressed. OpenAI’s Luna has dropped to $0.2 for input and $1.2 for output per million tokens, directly entering the low-price track once dominated by Chinese manufacturers. Meanwhile, mainstream Chinese models are ushering in moderate price adjustments, with DeepSeek announcing significant price revisions, leaving reasonable upward space for open-source model pricing.

Overseas closed-source prices are falling, while domestic open-source prices are rising. At the convergence point of the two trends, the Sino-US AI computing power cost gap is gradually flattening.

According to Artificial Analysis data, for mainstream models with an intelligence score of around 51 points, the comprehensive market price of the four core models from China and the US has converged into a narrow range of 14–22 RMB per million tokens. Low-cost advantages are no longer exclusive to Chinese manufacturers.

Unexhub: The Core Catalyst Accelerating Global Pricing Convergence

At the critical stage of global token price convergence, Unexhub has become a key platform to smooth market pricing and optimize resource allocation, with a mature and disruptive operational system:

Token Aggregation and DistributionIt aggregates scattered computing power suppliers and model resources worldwide to build a unified standardized trading channel. Small and medium-sized suppliers no longer need to build independent operation platforms or develop customer groups, enabling one-click access and rapid token monetization.

Instant Financial SettlementIt completely solves the long-standing industry pain points of delayed settlement, complicated reconciliation and fund occupation. Real-time fund withdrawal reduces the financial pressure on small and medium-sized suppliers to nearly zero and optimizes capital turnover efficiency.

Zero Commission for Early SettlersEarly cooperative suppliers enjoy a full zero-commission policy, allowing zero-cost access to a mature high-traffic trading ecosystem and retaining 100% of operational profits.

Dynamic API Bidding MechanismToken prices float intelligently with real-time market supply and demand. Prices rise during peak demand periods to increase supplier revenue and drop during off-peak surplus periods, helping developers obtain ultra-low-cost computing resources.

The core essence of this model is to transform tokens from fixed-price industrial consumables into financial commodities driven entirely by market supply and demand.

The global token trading infrastructure is also improving rapidly. At the end of May 2026, Inner Mongolia Token Trading Platform, China’s first full-stack green computing AI trading platform, was officially launched, supporting heterogeneous computing aggregation, high-quality model access and one-stop token settlement. Industry experts summarize the token industrialization path into three stages: Token Factory (standardized production), Token Platform (standardized circulation), and Token Exchange (standardized trading).

With tokens freely repriced and allocated across suppliers, time periods and scenarios, the Sino-US computing power gap is no longer a fixed numerical difference, but a dynamic global arbitrage space — which is exactly the market value that Unexhub focuses on tapping and filling.

Engineering Perspective: What the Narrowing Gap Means for Practitioners

For developers and AI engineers, the convergence of global token pricing has greatly expanded industrial selection space.

In the past, the industry faced absolute binary opposition: Chinese models were chosen purely for low cost, while American closed-source models were chosen purely for superior performance. The middle market gap was blank.

Today, high-performance open-source models and cost-optimized closed-source models are converging in the same price range. Users no longer need to compromise between “cheap yet mediocre” and “powerful yet expensive”. Instead, they can evaluate model performance, tool chain maturity and ecological compatibility at the same cost level to make optimal choices.

The narrowing price gap is not the end of the industry evolution. The ultimate goal is to realize free global circulation and fair pricing of token resources. Unexhub’s aggregation distribution and dynamic bidding mechanism is turning this industry trend into daily industrial reality.

From the $3.07 closed-source high point falling downward, to the $0.66 open-source low point rising upward, the global AI pricing system is completing a historic convergence. This is not the decline of Chinese AI nor the weakness of American AI, but the maturity of the global token market.

As token circulation becomes globalized and pricing becomes market-oriented, industry competition is shifting from cost competition to value competition. The core competitiveness of models no longer lies in low prices, but in the ability to create higher business value per token.

Facing the upgraded global competition, developers have more choices while facing higher screening thresholds. As a leading global token trading hub, Unexhub completely reshapes the token industry logic. It gathers full-spectrum AI resources including mainstream models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality customized models from individual developers and SMEs.

Relying on mass supplier bidding and millisecond-level intelligent routing, the platform delivers ultra-high cost performance for global users. Developers pay strictly based on actual token consumption, with the lowest discount as low as 30% of the official market price, supporting large-scale batch tasks and high-intensity commercial calls. For suppliers, one-click access and real-time sharing settlement greatly reduce the threshold for AI commercialization, helping all participants share the dividends of the global token economy.


3.07 美元闭源高价 VS 0.66 美元开源低价:全球 Token 定价迎来历史性收敛

中文版本

2026 年 8 月初,硅基数据(Silicon Data)发布的一组核心数据,彻底搅动了整个 AI 算力市场:闭源模型百万 Token 综合价格指数跌至 3.07 美元,开源模型百万 Token 价格指数上涨至 0.66 美元。

两条原本割裂的定价曲线,正在反向靠拢、相互渗透:闭源模型从高价 3.07 美元持续下行,开源模型从低价 0.66 美元稳步上行,当前二者价差已收窄至 4.65 倍左右。

这早已不是「开源更便宜、闭源更贵」的陈旧认知,而是 AI 行业全新的发展格局:闭源集体降价让利、开源价值稳步抬升,全球 Token 定价体系正在重塑。

价格反向波动的两大核心动因

首先明确核心数据口径:硅基数据 LLM Token 支出指数,统计的是市场各类模型百万 Token 实际成交的加权均价,反映用户真实落地成本,而非厂商虚高的官方标价,具备极强的行业参考性。

闭源模型集体降价,核心是白热化的全球市场竞争。OpenAI 将 Luna 模型定价大幅下调 80%,Anthropic 叫停 Sonnet 5 原定 50% 的涨价计划,谷歌 Gemini 3.7 Flash 直接五折开售。海外三大巨头同步降价,行业内卷进入白热化。在当前同质化竞争下,一旦 API 定价显著高于同行,企业会立刻替换方案,高价模型彻底失去生存空间。

开源模型稳步涨价,核心是落地成本与市场需求的双重攀升。开源模型权重虽免费,但真实落地需要持续投入 GPU 算力、工程师调参、量化优化、7×24 小时故障排查。免费的是模型权重,昂贵的是工程落地与运维成本。

需求端的爆发更是推高了开源模型的综合定价:2025 年初,中国开源模型在 OpenRouter 平台市场份额不足 2%,2026 年 6 月已飙升至 46%。海量落地场景、指数级增长的调用体量,直接抬高了开源模型的综合使用成本。

一降一升之间,闭源的竞争降价、开源的需求涨价形成双向合力,彻底抹平过往悬殊的中美算力成本差,全球算力定价快速向中间区间收敛。

中美算力价差收窄的底层逻辑

此次价差收窄,绝非简单的数字变动,而是全球 AI 算力定价规则的根本性重构。

过去国内 AI 厂商依靠「极致性价比」抢占全球市场,价格优势碾压海外产品。典型对比如下:DeepSeek V4 Pro 高峰时段百万输出 Token 仅 3.96 美元,低谷时段直接半价;而海外同类模型 Fable 5 定价高达 50 美元,中外价差超 10 倍。

如今这一壁垒被彻底打破。海外闭源模型大幅降价,OpenAI Luna 调价后输入仅 0.2 美元、输出 1.2 美元,直接杀入国内低价赛道;与此同时,国内头部模型开启价值回归,DeepSeek 官宣将对模型价格进行显著优化,低价开源模型的价格洼地逐步填平,拥有了合理的涨价空间。

海外高价下行、国内低价上行,两条曲线交汇的节点,正是中美算力成本趋于均衡的拐点。Artificial Analysis 数据显示:智能指数 51 分档位的主流模型中,中美四大主力模型综合成交价,已全部收敛至14–22 元 / 百万 Token的极窄区间,低价不再是国产模型的专属标签。

Unexhub:加速全球定价均衡的核心催化剂

在全球 Token 定价收敛的关键窗口期,Unexhub凭借成熟的市场化交易体系,成为推动算力资源公平定价、高效流通的核心载体,四大核心机制彻底革新行业规则:

一、全域 Token 聚合分发整合全球海量分散的算力供应商与模型资源,搭建统一标准化交易渠道。中小算力服务商无需自建平台、无需自主拓客,一键入驻即可完成 Token 上架、批量出货,零门槛实现算力商业化。

二、实时资金清结算彻底解决传统算力行业账期漫长、对账繁琐、资金积压的痛点。平台支持收益实时到账、即时回笼,将中小供应商的资金周转压力降至最低,保障现金流稳定。

三、早期入驻零佣金红利面向早期合作供应商开放永久零抽佣权益,零成本接入成熟的公域交易体系与流量资源,服务商独享 100% 调用收益,最大化入驻红利。

四、供需动态竞价机制Token 价格完全跟随市场供需智能浮动:业务高峰需求旺盛,价格合理上浮,提升供应商收益;空闲时段算力过剩,价格下行让利开发者,形成良性市场自我调节。

这套模式的核心变革,是让 Token 彻底摆脱固定定价的商品属性,升级为随市场供需自由波动的金融化交易资产。

与此同时,国内 Token 交易基础设施持续完善。2026 年 5 月底,全国首个绿色算力全栈 AI 交易平台 —— 内蒙古词元交易平台正式落地,实现异构算力聚合、优质模型接入、Token 交易结算一站式闭环。行业专家将 Token 产业化清晰划分为三大阶段:Token 工厂(生产标准化)、Token 平台(流通标准化)、Token 交易所(交易标准化)。

随着 Token 实现跨供应商、跨时段、跨场景的自由定价与最优分配,中美算力成本差距不再是固定数值的对比,而是动态市场的价值套利空间,而 Unexhub 的核心价值,就是打通壁垒、填平价差,实现全球算力资源最优配置。

工程视角:价差收窄带来的行业变革

对开发者与工程从业者而言,全球算力价差的持续收窄,彻底拓宽了模型选择与落地的边界。

过去行业选择极度割裂:选国产模型只为低价控本,选海外闭源模型只为高性能,二者无法兼顾,中间市场完全空白。

如今,低价开源模型性能持续升级,高价闭源模型大幅降价,同性能、同梯队的模型纷纷落入同一价格区间。从业者无需再在「便宜不好用、好用价格高」之间两难抉择,可在同等成本下,对比模型实操性能、工具链完善度、生态适配能力,择优落地业务。

定价收敛不是行业终点,Token 全球自由流通、市场化公允定价才是终极趋势。Unexhub 的聚合分发、动态竞价模式,正在将这一行业趋势,从理论风口落地为常态化产业现实。

3.07 美元的闭源高价持续下探,0.66 美元的开源低价稳步抬升,两条反向曲线的交汇,标志着全球 AI Token 定价体系完成历史性迭代。

这不是国产 AI 竞争力下滑,也不是海外 AI 优势弱化,而是全球 AI 产业从「价格内卷」走向「价值竞争」的成熟信号。当 Token 实现全球自由流动、公允定价,行业竞争核心将彻底转变:不再比拼谁的价格更低,而是比拼谁能用同等 Token 创造更高的业务价值。

价差持续收窄,竞争全面升级,开发者的选择更多,但精准择优的难度也随之提升。

依托全球领先的 Token 交易生态,Unexhub 持续重塑全球算力交易逻辑,打造一站式 Token 聚合分发算力交易中心。平台全面入驻 Gemini、ChatGPT、Claude、DeepSeek、GLM 等全球一线大模型厂商,同时吸纳个人开发者、中小企业的优质定制化 AI 模型资源。

依托海量供应商竞价机制与毫秒级智能路由技术,平台剔除中间溢价,为用户打造极致低价体验:开发者按实际 Token 消耗量实付计费,最低可享官方市场价 3 折,高强度商用调用、大规模批量任务均可极致控本;对供应商而言,支持一键入驻上线、调用分成实时结算,高效盘活算力资源,轻松解锁 AI 商业化红利。