Jason Says
Today's sharpest warning signal: GitHub Copilot switching to token-based billing, plus one company accidentally burning $500M on Claude in a single month — the 'runaway AI cost era' has arrived. Enterprises and solo developers must start taking AI usage governance seriously, or your invoices will teach you the hard way.
GitHub Copilot announced a switch from flat subscription pricing to token-based usage billing, triggering fierce backlash from the developer community. The move signals the end of the 'free lunch' era for AI coding tools. As both user counts and usage volumes explode, platforms are inevitably passing costs back to users — heavy users will feel the pain most.
An undisclosed company failed to set usage caps on employee Claude licenses, resulting in a $500M API bill in a single month. The incident highlights two realities: Anthropic's Opus models consume tokens aggressively, and enterprise AI cost governance remains largely nonexistent. It's a stark warning for any team deploying Claude at scale internally.
EveryInc open-sourced the Compound Engineering Plugin, officially supporting Claude Code, Codex, Cursor, and other leading AI coding tools. Its core philosophy — each engineering task should make the next one easier — uses reusable AI Skills to combat technical debt accumulation. It's a rare plugin in the Claude Code ecosystem focused on compounding engineering value.
TechCrunch tested Google's new Gemini Spark, a 24/7 AI assistant designed to automate daily tasks like inbox summarization and local event planning. Hands-on experience proved genuinely useful, but the central question remains: why did Google build it as a standalone product instead of integrating it directly into the Gemini app? Product strategy confusion is showing.
OpenAI officially unveiled the Rosalind Biodefense program, granting vetted developers and US government partners access to a dedicated GPT-Rosalind model focused on biodefense, public health, and pandemic preparedness. This marks OpenAI's first model tailored specifically for national security use cases, deepening AI's penetration into high-stakes government applications.
Meta is reportedly building an AI-driven wearable pendant, marking another hardware bet following the Ray-Ban smart glasses. The AI hardware battlefield has officially expanded from phone accessories to intimate wearables worn on the body. Humane AI Pin's failure clearly hasn't dampened big-tech enthusiasm for the category.
Prominent AI investor Bindu Reddy shared data showing open-source AI token consumption growing exponentially, approaching the volume of Google's Gemini models. Kimi, DeepSeek, and GLM already handle 50% of common tasks adequately. For SaaS developers relying on closed-source APIs, the cost signal is clear — the window to migrate toward open-source models is opening.
CONF-KV 提出用模型每步解码时的「下一 token 预测置信度」来动态决定保留多少 KV Cache，高置信时激进压缩、低置信时保留更多上下文。对开发者的意义：在不换模型的前提下，长文档处理的 GPU 显存和推理成本可大幅降低，部署 10 万 token 以上长文本应用的工程成本有望实质性下降。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free