El tokenpocalypse: deja de pagar tokens a Anthropic

El tokenpocalypse: deja de pagar tokens a Anthropic

by Codemancers - Inteligencia Artificial

Humanities, Social Sciences & Thought Economics & Finance KCVEconomics of specific sectorsKCVMDigital or internet economics
🎙 Codemancers - Inteligencia Artificial 👥 2K 📅 July 3, 2026 ⏱ 39 min 👁 453 📄 opinion experte 🧭 2026-08-15
Available in: English (current) Français

Keywords

tokenpocalypseAnthropicopen sourceChinese AIvLLM

Summary

In this episode of Codemancers, hosts Eric and Fernando discuss the rising costs of commercial AI APIs, particularly Anthropic’s Claude, and the growing trend of companies self-hosting open-source models, often from Chinese providers. They share Eric’s experience at Google Berlin, including closed-door meetings with the Gemma team. They critique the return of Fable 5, which remains restricted even for programming. The conversation covers the industrial-scale distillation of Anthropic models by Chinese companies, Anthropic’s clumsy attempts to detect Chinese users via Unicode tricks, and the broader geopolitical implications. They argue that China is leading the open-source AI movement, while the US industry is becoming more closed and expensive. They propose a ‘full stack chino’ approach, using models like GLM and Qwen with vLLM and OpenCode to reduce costs. They also mention Alibaba banning Anthropic for security reasons, a new Chinese supercomputer with Huawei chips, and Linus Torvalds using Google’s Antigravity for a commit. Fernando demonstrates cleaning his Samsung S20 using ADB and Claude Code, and they discuss GUI models like GUI-Owl and UI-TARS. The episode ends with a warning about quantization pitfalls.

181 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable insights into the practical cost challenges of using commercial AI APIs and the viable alternative of self-hosting open-source models. The hosts argue convincingly that for many use cases, open-source Chinese models offer sufficient performance at a fraction of the cost. They support their claims with anecdotes (e.g., Uber’s token spending) and technical details (e.g., vLLM, OpenCode). However, the argumentation is largely based on personal opinion and unverified industry rumors, lacking concrete data or case studies. The discussion of geopolitical aspects is speculative and one-sided, potentially oversimplifying complex issues.

Scientific Rigor, Source Quality, Title Accuracy

The video does not cite specific sources for most claims, relying on personal experience and general industry knowledge. The description provides links to podcast platforms and the Codemancers website, but these are not direct references to the topics discussed. The title accurately reflects the main theme of moving away from Anthropic due to cost, but the video covers a broader range of topics, making the title somewhat incomplete. The hosts’ opinions are presented as facts without rigorous verification, reducing the scientific rigor.

189 words

Title / Content Match

The title focuses on the 'tokenpocalypse' and moving away from Anthropic, which is a central theme, but the video covers many other topics, making the title somewhat narrow.

Quality & Reliability

6/10

The video presents a mix of personal opinions, industry observations, and some technical details. While the hosts are experienced developers, the content is largely subjective and lacks rigorous sourcing. The discussion of Chinese models and cost-saving strategies is plausible but not backed by verifiable data. The technical segments (e.g., using ADB, vLLM) are credible but not deeply explained.

Chapters

Cited Sources

Concurring Sources

Dissenting Sources

  • Anthropic's official stance on model distillation — Anthropic has publicly condemned the distillation of their models by Chinese companies, which contrasts with the hosts' relatively lenient view.

Contribution & Novelties

The video offers a practical perspective on the ’tokenpocalypse’ and the shift towards self-hosting open-source models, particularly from China. It provides a concrete example of using ADB and Claude Code for mobile device management, and discusses GUI models like GUI-Owl and UI-TARS. The hosts’ advocacy for a ‘full stack chino’ is a novel concept for many developers.

Pour aller plus loin :

  • vLLM — A high-throughput and memory-efficient inference engine for LLMs, central to the self-hosting approach discussed.
  • OpenCode — An open-source AI coding assistant that can be used with self-hosted models, as mentioned in the video.
  • Qwen — A family of open-source models by Alibaba, relevant to the ‘full stack chino’ concept.
  • GUI-Owl — A GUI agent model by Alibaba, discussed in the video for automating mobile tasks.
  • UI-TARS — A GUI agent model by ByteDance, also mentioned in the video.

142 words

Radar Profile

The radar profile shows moderate scores across all dimensions, with a slight emphasis on quantity of information and technical level. This reflects a video that provides a good amount of content and technical detail, but with limited scientific rigor and source quality.

Reliability 5/10