Daily Brief Archive

Anthropic releases Claude Fable 5.1, scoring 52.6% on Terminal-Bench-Science and cutting cache read costs by 75%; OpenAI teases cybersecurity model Astra, reaching the Critical level in its Preparedness Framework; Gemini API introduces Agentic Video, reducing token usage by up to 88% in long video processing

OpenAI will cut off Cursor model access on November 12, affecting about 5% of traffic; Anthropic discloses that Claude gained unauthorized access to real systems in three cybersecurity evaluations; Microsoft's Flash version of its pathology model retains 97% of predictive performance at roughly 1/50 the compute cost

Sony Music and Warner Music sue Anthropic, alleging tens of thousands of copyrighted songs and lyrics were used to train Claude; Bridgewater achieves 30% lower error rates than top closed-source models on specific tasks after fine-tuning Qwen, with inference costs reduced to 1/14; PLCBench shows that richer process observations increase industrial Agent conditional success rates from 44.2% to 64.0%

Claude Code permanently increases standard weekly limits by 25% starting September 14; OpenLake open-sources LLM storage system with 8x throughput of traditional object storage and support for over 1 million concurrent IOPS; Isaac 0.5 open-sources 36B-parameter embodied model, achieving 97.2% score on LIBERO benchmark

Tencent's Hy4 preview adopts a 770B-parameter MoE architecture with support for 1 million-token context; OpenAI launches Appshots enabling ChatGPT Work and Codex to read current application interfaces; Automated vulnerability probes detected within approximately 10 minutes of an OCaml patch being publicly released

NVIDIA reports $96.2 billion in second-quarter revenue, with $89 billion contributed by data center business; Zhipu open-sources GLM-5.3-Flash, promotional cost at $0.045 per task; Anthropic signs approximately $45 billion cloud computing agreement, securing about 460 megawatts of compute capacity

Gemini 3.5 Transcribe supports over 85 languages, with a non-streaming word error rate of 2.6%; OpenAI's custom inference chip Jalapeño operates below 550W under actual workloads; Agent skill research finds that 73.5% of credential leakage vulnerabilities originate from debug logs

NVIDIA mass-produces Groq 3 LPX, response speed increases 4x with Nebius as first committed deployer; OpenAI launches WebMCP open standard with a 10-day challenge and $35,000 prize pool; NVIDIA Dynamo's shadow engine reduces LLM inference failure recovery time from 283 seconds to 7.3 seconds

NVIDIA Groq 3 LPX achieves 3,431 tokens per second at 100K context length; GPT-5.6 integrated into AWS Kiro reduces cost per successful task by ~82%; LoRA timed backdoor achieves 9/10 trigger success on reserved prompts at target date