Daily Brief Archive

OpenAI releases GPT-5.6 trio, Sol scores 53.6 on long-task benchmark, outperforming Claude Fable 5 by 13.1 points at roughly one-quarter the cost; OpenAI launches ChatGPT Work autonomous agent and merges Codex, desktop app now globally available; SambaNova raises $1 billion, reaches $11 billion valuation to accelerate inference chip deployment

OpenAI launches full-duplex voice model GPT-Live, capable of background delegation to GPT-5.5 and supporting real-time translation; Cursor collaborates with SpaceXAI to release Grok 4.5, achieving Opus-level performance while being faster and cheaper; OpenAI audits SWE-Bench Pro, finds 30% of tasks corrupted, and retracts prior benchmark recommendation

Meta releases first 「reason-then-generate」 agent-style image model Muse Image; Apple extends partnership with Broadcom to 2031 to co-develop custom AI server chips; Fable 5 orchestration mode achieves 96% performance at 46% cost

Anthropic discovers Claude's internal 「global workspace」J-space, resembling human consciousness, enabling access to hidden thoughts and use in security monitoring; Tencent open-sources 295B MoE model Hy3, matching 2–5x larger flagship models with only 21B activated parameters and reducing hallucination rate to 5.4%; first fully autonomous LLM-powered ransomware JadePuffer exposed, encrypting 1,342 configurations without human intervention

Meituan open-sources 1.6 trillion-parameter LongCat-2.0, trained entirely on 50,000 domestic AI chips without NVIDIA, achieving second-best Agent capability globally; Meta shifts to selling GPU compute and opens AI infrastructure to external clients, driving stock up nearly 9%; Doubao and Qwen discontinue Agent features on July 15 due to new anthropomorphic AI regulations, while Alibaba bans Claude, Meituan restricts Doubao, and major firms collectively pivot to in-house models

Mistral's open-source math model Leanstral 1.5 solves 587/672 problems on PutnamBench; malicious AI agent skill bypasses security checks affecting 26,000 users; NVIDIA's HORIZON framework achieves 100% pass rate on RTL chip design benchmarks

Meta's 「Watermelon」 model matches GPT-5.5 but Zuckerberg says AI progress falls short of expectations; Sysdig discovers first end-to-end ransomware attack executed by an AI agent exploiting a Langflow RCE vulnerability; Alibaba DAMO Academy AI discovers 4 entirely new superconductors using just 28 GPU hours

Meta plans to launch an AI cloud service to sell excess compute, directly challenging AWS, Azure, and Google Cloud; OpenAI proposes transferring 5% equity to the U.S. government to establish a public wealth fund; Microsoft Azure CLI suffers 81 million password spraying attempts, breaching 78 accounts across 64 organizations

OpenAI will launch its first hardware product, the Codex mechanical keyboard with 13 mechanical switches, on July 15; The U.S. Department of Commerce lifts export controls on Anthropic's Fable 5 and Mythos 5, with Fable regaining the top spot on Cursor; NVIDIA open-sources the ASPIRE robotics skill library, boosting success rates on unseen tasks from zero to 31%