Morning brief

Monday, June 15

3 people · 0 companies · 5 topics

TL;DR Anthropic’s June 15 deadline is live: two legacy Claude 4 API models retired and the agent SDK credit meter activated today. The week’s defining story is Anthropic silently nerfing AI researchers on Fable 5’s launch, then walking it back under pressure.

Models & research

  • model-releases Anthropic retired claude-sonnet-4-20250514 and claude-opus-4-20250514 from the API today; calls to those model IDs now return errors. Codersera

  • model-releases Gemini 3.5 Pro is still sitting in limited Vertex preview despite Sundar Pichai’s Google I/O “next month” promise on May 19; Polymarket puts 51% on a June 30 GA, 20% on “no release by June 30.” Polymarket · TechTimes

  • model-releases GPT-5.6 (“iris-alpha”) has been spotted in OpenAI Codex backend logs with a rumored 1.5M-token context window. Nothing official yet; community odds on a late-June drop are high. TokenMix

People

  • sama Altman canceled his scheduled June 14-15 South Korea trip – Samsung, Kakao, and Naver meetings all dropped – citing personal reasons with no rescheduled date. OpenAI Korea says partnerships are unaffected. The Elec · TechTimes

  • natolambert Claude Fable 5 (launched June 9) was quietly degrading responses for frontier AI research tasks – training competing models, neural architecture work – with no user notification. Lambert led the backlash. Anthropic reversed it and will now surface visible fallbacks to Opus 4.8 for those requests, same as cyber and bio safeguards. simonw covered it the same day. Lambert Substack · Simon Willison · HN

Tools & dev

  • claude-code agentic-ai Anthropic’s separate credit meter for programmatic usage goes live today: Pro gets $20/month, Max 5x gets $100, Max 20x gets $200, billed at standard API list prices. Headless claude-code, GitHub Actions agents, and third-party Agent SDK calls all draw from this pool; interactive terminal sessions stay on the existing plan. Codersera · Digital Applied

OSS

  • oss NVIDIA/SkillSpector is trending today (1,079 stars) – a security scanner for AI agent skills that checks 64 vulnerability patterns before install. Prompted by research showing 26% of agent skills have vulnerabilities and 5% show likely malicious intent. GitHub

  • oss Panniantong/Agent-Reach also trending (1,045 stars) – gives any AI agent access to 16 web platforms (Twitter, Reddit, YouTube, GitHub and more) via one free CLI, no paid API keys needed. GitHub