1989 年,鲍里斯·叶利钦在休斯顿的一家 Randalls 超市前停下脚步,被琳琅满目的冰淇淋震惊了。1 OpenRouter 就是 AI 领域的那个超市货架。
而消费者的选择出人意料:OpenAI 一年前发布的开源模型 GPT-OSS 120b2 占据了 Anthropic 的 Opus 4.8 流量的 36%。3
为什么一个 2025 年 8 月的模型,仍然占据着一个几周前才发布的前沿模型三分之一的流量?
Token 市场已经细分了。
细分之所以发生,是因为买家需求各不相同。
这种细分正在竞争驱动下加速。上周 Anthropic 发布了 Opus 5,比 Fable 更小、更便宜,4 明确是为了争夺月之暗面 Kimi 3 所瞄准的市场。5 在中端模型市场,Poolside 推出了 Laguna S 2.1,一款美国中端市场模型。6
规模(小、中、大、XL)、来源(美国 vs 中国)、架构(稠密 vs 稀疏)、精度(专注编程还是通用)、速度(每秒 token 数)、模态(纯文本还是视觉);AI 有许许多多不同的风味。
我花了一个周末替换运行我的智能体的模型。现任是 Gemma 4 26b;挑战者是 Laguna S 2.1,一个 1180 亿参数的模型。按照我预期会起作用的每一项数字,这个 118b 模型本应更慢。
在我的 M5 Max 上,两者生成速度相同,因为 Laguna 是混合专家架构:1180 亿参数驻留在内存中,但每个 token 只激活 80 亿。一个 118b 模型现在以 26b 模型的解码成本运行,这把前沿级质量拉低到了本地层级。
准确率的提升体现在关键之处。我的本地技术栈运行一个编码与邮件智能体进行工具调用,而在我轮换使用过的各个模型中,随着激活参数上升,工具调用失败率从 29.4% 降至 20.1%。7 Laguna 相比它所替换的 26b 模型将错误率降低了 7 个百分点。
分层是一个健康竞争市场的标志。前沿仍然服务于世界上最难的 token。它不再需要服务所有 token,而我笔记本电脑上的层级刚刚获得了一个高得多的上限;这一趋势将被竞争不可阻挡地向前推进。
-
鲍里斯·叶利钦 1989 年造访休斯顿一家杂货店,1989 年 9 月 16 日,Clear Lake 的一家 Randalls。↩︎
-
推出 gpt-oss(OpenAI),发布于 2025 年 8 月 5 日。↩︎
-
OpenRouter 模型活动页面,7 天平均值,检索于 2026-07-27:GPT-OSS-120b、GLM 5.2、Claude Opus 4.8。↩︎
-
Anthropic 以一半价格推出 Claude Opus 5,发布于 2026 年 7 月 24 日。↩︎
-
月之暗面的 Kimi 3 有望缩小与 Anthropic 的 Opus 4.8 之间的差距(TechCrunch)。↩︎
-
推出 Laguna S 2.1(Poolside),一个总参数 118B、激活参数 8B 的开放权重模型,发布于 2026 年 7 月 22 日。↩︎
-
作者的生产数据:来自一个本地编码与邮件智能体的 MCP 工具调用日志,在五个月内跨三个本地模型测量。↩︎
In 1989 Boris Yeltsin stopped at a Randalls supermarket in Houston, stunned by the variety of ice cream.1 OpenRouter is that supermarket aisle for AI.
And shoppers make surprising choices : OpenAI’s year-old open-source model GPT-OSS 120b2 commands 36% of Anthropic’s Opus 4.8 volume.3
Why does a model from August 2025 still hold a third of the traffic of a frontier model that shipped weeks ago?
The market for tokens has segmented.
Segmentation happens because buyer needs vary.
The segmentation is accelerating driven by competition. Last week Anthropic shipped Opus 5, smaller & cheaper than Fable,4 explicitly to contest the ground that Moonshot’s Kimi 3 targets.5 In the mid-model-market, Poolside launched Laguna S 2.1, a US mid-market model.6
Size (small, medium, large, XL), origin (US v China), architecture (dense vs sparse), accuracy (coding focused or general), speed (tokens per second), modality (text-only or vision) ; there are many flavors of AI.
I spent the weekend replacing the model that runs my agent. The incumbent is Gemma 4 26b; the challenger is Laguna S 2.1, a 118-billion-parameter model. By every number I expected to matter, the 118b model should have been slower.
On my M5 Max both generate at the same speed, because Laguna is a mixture-of-experts architecture: 118 billion parameters live in memory, but only 8 billion activate per token. A 118b model now runs at the decode cost of a 26b model, which pulls frontier-class quality down into the local tier.
The accuracy shows up where it matters. My local stack runs a coding & email agent on tool calls, & across the models I have cycled through, the tool-call failure rate falls from 29.4% to 20.1% as active parameters climb.7 Laguna reduces error rates by 7 percentage points over the 26b model it replaced.
Segmentation is the sign of a healthy competitive market. The frontier still serves the world’s hardest tokens. It no longer has to serve all of them, & the tier on my laptop just got a much higher ceiling ; a trend that competition will push forward inexorably.
-
Boris Yeltsin’s 1989 visit to a Houston grocery store, a Randalls in Clear Lake, September 16, 1989. ↩︎
-
Introducing gpt-oss (OpenAI), released August 5, 2025. ↩︎
-
OpenRouter model activity pages, 7-day average, retrieved 2026-07-27: GPT-OSS-120b, GLM 5.2, Claude Opus 4.8. ↩︎
-
Anthropic debuts Claude Opus 5 at half the price, launched July 24, 2026. ↩︎
-
Moonshot’s Kimi 3 is expected to close the gap with Anthropic’s Opus 4.8 (TechCrunch). ↩︎
-
Introducing Laguna S 2.1 (Poolside), a 118B-total, 8B-active open-weight model released July 22, 2026. ↩︎
-
Author’s production data: MCP tool-call logs from a local coding & email agent, measured across three local models over five months. ↩︎