在人工智能领域，计费与核算的基本单位是“Token”（通常约为四个字符的文本片段）。然而与传统的标准化度量单位不同，Token在实际应用中具有高度不确定性与极大的定价差异：前沿任务消耗的Token数量难以提前精确预估，且不同开发商的收费标准悬殊，例如中国初创企业深度求索（DeepSeek）的主流模型每Token价格仅为美国顶尖实验室Anthropic前沿模型的约四十分之一（1/40）。这种计费的复杂性与不可预测性常常导致企业面临意外的高额账单，进而催生了一种新型商业生态——AI模型聚合交易平台（Model Marketplace）。
在这一新兴赛道中，最具代表性的是由亚历克斯·阿塔拉于2023年联合创立的OpenRouter。该平台允许开发者和企业通过统一接口在不同模型之间自由无缝切换，既能在服务宕机时实现故障转移，也能在特定厂商降价促销时智能套利。尽管亚马逊的Bedrock和谷歌的Vertex AI等传统云计算巨头也提供类似服务，但中立的独立平台更受开发者青睐。得益于业务的爆发式增长，全球支付巨头Stripe在今年8月同意斥资高达75亿美元收购成立仅三年的OpenRouter，凸显了中立模型路由器在AI产业版图中的战略价值。
OpenRouter最初因打破模型实验室对基准测试的过度刷分（“benchmaxxing”）并提供真实消费排行动向而声名鹊起，如今更已成为企业控制推理成本的关键中枢。在平台上，以DeepSeek以及Z.ai和小米为代表的高性价比中国模型主导了庞大的Token消耗量。阿塔拉建议企业应根据开放程度将任务划分为低、中、高“熵”层级：将代码分类路由或常规审查等低中熵任务分派给低成本模型处理，而将从零编写复杂代码等高熵任务保留给昂贵的前沿旗舰模型，以此避免不加拆分所导致的巨额算力资金浪费。
The artificial intelligence ecosystem is anchored to a unique unit of account: the token, typically representing roughly four characters of text. Unlike traditional measuring units, tokens exhibit extreme price volatility and consumption unpredictability; enterprise tasks can consume wildly variable token volumes, while vendor pricing diverges dramatically—with Chinese startup DeepSeek pricing tokens at approximately one-fortieth (1/40th) the rate of cutting-edge models from leading American labs like Anthropic. This opacity and risk of runaway costs have catalyzed the rapid emergence of model marketplaces designed to optimize inference spend and workflow routing.
Leading this nascent industry is OpenRouter, co-founded in 2023 by Alex Atallah to counter potential single-provider monopolies by enabling dynamic routing across diverse foundation models. OpenRouter facilitates automated failover during outages and allows real-time arbitrage when competing providers discount token rates, charging a nominal routing fee per transaction. While hyperscalers like Amazon (via Bedrock) and Google (via Vertex AI) operate competing marketplaces, OpenRouter's vendor neutrality provides critical independence from model-producing cloud incumbents. Underscoring this immense strategic value, fintech giant Stripe agreed in August to acquire the three-year-old startup for a reported valuation of $7.5bn.
Initially popular among developers seeking objective performance evaluation beyond synthetic benchmarks ("benchmaxxing"), OpenRouter has evolved into an essential enterprise cost-containment hub. Demonstrating aggressive price sensitivity, cost-effective architectures from DeepSeek alongside Chinese providers Z.ai and Xiaomi dominate aggregate token consumption on the platform. To eliminate corporate waste, Atallah advocates decomposing workloads by "entropy": allocating low- and medium-entropy tasks, such as routing and structured code reviews, to budget-friendly models, while reserving premium frontier models exclusively for high-entropy generative initiatives like architecting code from scratch.
Source: The rise of the model marketplace
Subtitle: Businesses such as OpenRouter are thriving as companies try to cut costs
Dateline: Oct 8th 2026\n