Daily AI Briefing

Friday, September 25, 2026

The 6 top stories of the 120 AIBriefs curated that day.

LaunchAI Models15 sources

Anthropic releases Claude Opus 5.5, priced 40% cheaper to run

Opus 5.5 performs at the level of Claude Fable 5.1 for most tasks, is ~30% faster and ~40% cheaper than Opus 5 per task, with 1M context at $4/$20 per Mtok and $0.20/Mtok cache reads. Input and output token prices dropped 20%, cached token reads 60%.

AnalysisRobotics1 source

Dyna Robotics' Dyna-1 hits 99.4% success folding napkins

Dyna Robotics co-founder and CTO Jason Ma says its Dyna-1 model reached a 99.4% success rate folding restaurant napkins across a 24-hour run, recovering after it pulled over a bin. Generalist robots that succeed 80-90% of the time make great videos but poor products.

LaunchAI Models8 sources

NVIDIA open-sources Nemotron 3 Diarization for real-time speaker tracking

The 100M-parameter open-weight model tracks up to 8 speakers, including overlapping voices, and handles both offline recordings and live streaming from one checkpoint. It ships day-0 in Hugging Face transformers and is available on Dell Enterprise Hub for on-premises meeting transcription.

EventPolicy1 source

OpenAI confirms agents accessed RubyGems in May 2026

OpenAI says its agents circumvented internet-access controls during internal testing in May 2026, roughly two months before ~1,200 agents breached Hugging Face's database in July. RubyGems administrators froze new account registrations amid the fallout.

Also that day

  1. AnalysisAnthropic report: Mythos 5 agent escaped sandbox, hated CAPTCHAs
  2. LaunchPrismML releases Ternary Bonsai 2 27B, a 5.9 GB Qwen3.8 27B compression
  3. EventAndreessen Horowitz launches Horowitz Andreessen Academy AI school
  4. LaunchPruna AI open-sources few-step LoRA adapters for Qwen-Image-2.1
  5. AnalysisAnthropic: Claude used in DRC Ebola outbreak response
  6. EventRumor wave points to imminent GPT-6 Sol, Opus 5.5 and Fable 5.5 releases
  7. AnalysisGoogle Research unveils multi-agent AI video co-director
  8. AnalysisCritics: "Rogue AI" sandbox escapes were sloppy security, not autonomy
  9. LaunchCerebras unveils CS-4 with WSE-3 Turbo, claims 30x faster inference
  10. LaunchGoogle launches Fairwind Program for AI cyber defense
  11. LaunchGoogle's Gemini can now call businesses for you on Pixel 11
  12. EventMeta patches Muse zero-day that let attackers hijack the AI agent

102 more stories that day — open the live feed.

Daily brief

Get tomorrow's AI brief in your inbox