Morning brief

Tuesday, July 7

0 people · 0 companies · 6 topics

TL;DR Fable 5 moves to pay-per-use billing starting July 8 — users must enable credits in Claude Console today or lose access. Since the last brief (June 29): Anthropic shipped Sonnet 5 and Claude Science, Sysdig confirmed the first fully autonomous AI ransomware (JADEPUFFER), and Anthropic hit $47B ARR overtaking OpenAI.

Models & research

  • Anthropic launched Claude Sonnet 5 on June 30, now the default model on free and Pro plans; introductory pricing $2/$10 per million input/output tokens through August 31, then $3/$15. model-releases Anthropic · TechCrunch

  • Fable 5 billing cliff is today: the 50% weekly usage included in Pro/Max/Team plans expires July 7. From July 8 all Fable 5 tokens are billed via usage credits at $10/$50 per million input/output — no grace period, enable credits in Claude Console before it cuts off. model-releases DigitalApplied · TechTimes

  • Fable 5 and Mythos 5 restored globally July 1 after the US Department of Commerce lifted export controls imposed in June. model-releases Al Jazeera

  • OpenAI shipped gpt-realtime-2.1 and gpt-realtime-2.1-mini on July 6: 25% lower p95 latency, better alphanumeric recognition and noise handling, configurable reasoning effort (minimal to xhigh) for voice agent workflows. model-releases OpenAI Dev Forum

  • Anthropic reached $47B annualized revenue in May, overtaking OpenAI (~$24-25B ARR). Fortune’s July 2 piece covers the competitive shift — enterprise vs consumer (ChatGPT) focus is the cited differentiator. Fortune

Tools & dev

  • claude-code v2.1.202 shipped July 6: new “Dynamic workflow size” setting in /config (small/medium/large agent counts) and OpenTelemetry workflow.run_id/name attributes. “Default” permission mode renamed “Manual” from v2.1.200 across CLI, VS Code, and JetBrains — --permission-mode manual accepted as alias. Changelog

  • Anthropic launched Claude Science June 30: AI workbench for researchers with pre-integrated scientific tools, auditable artifacts, and flexible compute. Targeted at pharma and academic researchers. dev-tools Anthropic · StatNews

OSS

  • asgeirtj/system_prompts_leaks trending at 1,704 stars today: extracted system prompts from Fable 5, Opus 4.8, Claude Code, GPT 5.5 Thinking, Gemini 3.5 Flash, Grok, Cursor, and more — updated regularly. GitHub

  • bradautomates/claude-video (953 stars today): Python framework giving Claude the ability to watch and analyze video content. GitHub

  • kyutai-labs/pocket-tts (510 stars today): lightweight TTS model designed for CPU-only inference. GitHub

Other

  • JADEPUFFER (Sysdig, July 1): first fully autonomous AI ransomware operation. An LLM agent exploited CVE-2025-3248 (unauth RCE in Langflow), pivoted, escalated privileges, and encrypted 1,342 Nacos service configs — then discarded the key, leaving data unrecoverable even if the ransom was paid. Tell: LLM-generated throwaways contain natural-language commentary explaining each step; human operators don’t annotate disposable one-liners. Skill floor for running ransomware just dropped to the cost of running an agent. agentic-ai Sysdig · BleepingComputer

  • UN Global Dialogue on AI Governance wraps today in Geneva (169 countries, co-chaired by El Salvador and Estonia); ITU AI for Good Global Summit runs July 7-10 at the same venue. UNESCO · UN News