# 为降本，亚马逊工程师正蒸馏Anthropic模型

- 来源：The Decoder：AI News（RSS）
- 作者：Matthias Bastian
- 发布时间：2026-06-30 02:05
- AIHOT 分数：50
- AIHOT 链接：https://aihot.virxact.com/items/cmqzj5qcp002uslqvg5mlomcz
- 原文链接：https://the-decoder.com/amazon-engineers-are-reportedly-distilling-anthropic-models-to-cut-costs-before-new-token-based-pricing-kicks-in

## AI 摘要

据The Information报道，担心成本上升，亚马逊工程师正蒸馏Anthropic模型，构建更小更便宜的内部版本。蒸馏由小型模型学习大型模型输出实现，亚马逊据称拥有相关权限（类似苹果与Gemini）。但Bedrock平台蒸馏服务不支持Claude，仅支持自家Nova和Meta的Llama。此举源于合作重新谈判：明年代Amazon将按处理token而非计算小时为Anthropic模型付费，可能导致成本飙升。亚马逊发言人否认成本提高，Anthropic强调性价比。此外今年亚马逊已向Anthropic追加250亿美元投资，并向OpenAI投资500亿美元。

## 正文

Amazon engineers are reportedly distilling Anthropic models to cut costs before new token-based pricing kicks in

Matthias Bastian View the LinkedIn Profile of Matthias Bastian

Jun 29, 2026

Worried about rising costs, some Amazon engineers are already distilling Anthropic models to build smaller, cheaper versions for internal use. That's according to a report from The Information. Distillation works by having a smaller model learn from a larger model's outputs. Amazon has certain rights to use Anthropic's models for this purpose, according to a person familiar with the matter, similar to Apple's arrangement with Google Gemini. Amazon does offer a distillation service on its Bedrock cloud platform, but Anthropic's Claude models aren't available there; only Amazon's own Nova models and Meta's Llama models are supported.

The effort ties back to a renegotiation of the partnership, according to The Information. Starting next year, Amazon will pay for Anthropic's models based on tokens processed rather than compute hours, which could push costs up sharply. An Amazon spokesperson pushed back, saying the changes from the expanded partnership won't raise costs. Anthropic points to lower prices relative to the performance its models deliver.

Amazon is reportedly exploring alternatives like OpenAI and its own Nova models. The company has invested up to $25 billion more in Anthropic and up to $50 billion in OpenAI this year.