Models11月12日 01:22

盲测显示 DeepSeek 在 95% 智能体任务中比肩 Carbon/Opus，价格低 40 倍

Bindu Reddy@bindureddy

Post

Why we picked it

做智能体开发的朋友可以看看这份盲测：DeepSeek 在 95% 的循环任务里打平 Carbon/Opus，但价格只有 1/40。如果你在跑短循环、高频调用的任务，换它能省不少钱；长循环场景还是要留个心眼，退化比较明显。

开发者 Bindu Reddy 发布盲测结果称，DeepSeek 在智能体循环任务中与 Carbon/Opus 达到 95% 的匹配度，而价格仅为其 1/40。这意味着在绝大多数常规智能体工作流中，选择 DeepSeek 可以大幅降低调用成本。

Body

根据该测试观察，DeepSeek 的短板出现在很长的智能体循环中：随着循环变长，其表现退化速度快于前沿模型。作者认为问题出在上下文压缩（compaction）环节，认为修复压缩机制后 DeepSeek 就能与这两款模型持平。

对开发者而言，这一结果指向一个务实的选型策略：短循环、高频调用的智能体任务可优先用低成本模型，长循环复杂任务仍需前沿模型兜底，混合路由可能是最优解。

Original · Bindu Reddy

In blind tests DeepSeek matches Carbon / Opus on 95% of agentic loops

It’s 40x cheaper and is way more attractive.

The only draw back is very long agentic loops where it seems to degrade faster than frontier models

Fixing compaction should bring it on par with them!

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include DeepSeek, 智能体, 推理模型, 开源模型.

Source attribution: curated from Bindu Reddy (@bindureddy).

#DeepSeek #智能体 #推理模型 #开源模型

Keep reading

Recent stories with the same tag or section

Papers BuildersSelected01:28

Study finds LLM moral choices lack stable value rankings

Averaged benchmark scores can manufacture a coherent moral hierarchy that no individual model inference actually exhibits, so alignment scores may overstate real consistency. Teams building safety evaluations or compliance checks for deployed models should replicate this pairwise-probability method before trusting aggregate scores. The inclusion of GPT-5.2, Gemini, and DeepSeek makes the finding directly relevant to cross-lab model selection.

#GPT-5.2 #Gemini #DeepSeek #AI safety

rohanpaul_ai@rohanpaul_aiSource

Industry US vs ChinaMulti-source00:23

Chinese contentNikkei：DeepSeek 等中国 AI 公司收入仅为 OpenAI、Anthropic 的 10%

日经这组数据挺有意思：DeepSeek 它们收入只有 OpenAI、Anthropic 的 10%，居然还能盈利，靠的是把算力成本压得很低。想搞清楚中美 AI 公司商业化路数差异的，值得算算这笔账。

#DeepSeek #OpenAI #Anthropic #行业收入

Event

Nikkei Asia@NikkeiAsiaSource

How-to BuildersMulti-sourceSelected21:51

AI Gateway: Why Teams Extract Model Calls Into One Layer

The non-obvious insight is that gateway adoption is driven less by convenience than by switching costs: without one, changing model providers means rewriting adapters and fallback logic across the codebase. Backend and platform engineers building multi-model applications should evaluate gateways before adding their second provider. The examples span OpenAI, Anthropic, Gemini, Qwen and DeepSeek, which maps directly onto the provider mix most teams in China and Southeast Asia juggle when balancing cost and capability.

#AI Gateway #OpenAI #Anthropic #DeepSeek

Event