Here is today's AI news. Story 1. OpenAI. OpenAI launched GPT-6 Astra, a new frontier model focused on cybersecurity and computer skills, and positioned it as its most advanced and aligned release yet.. OpenAI launched GPT‑6 Astra on September 3, 2026, calling it its most capable and aligned model. It sets new benchmarks for computer use, browsing, software engineering and cybersecurity, scoring 100% on ExploitBench versus 78.5% for GPT‑5.6 Sol. Astra is OpenAI’s first model to reach its “Critical” cyber threshold, capable of finding zero-day flaws and developing exploits. Access is initially restricted through Daybreak, with guardrails limiting offensive cyber tasks for broader users. Story 2. Google DeepMind. Google released Gemini 3.8 Flash, plus a restricted Cyber variant, continuing its rapid Flash-model cadence with stronger reasoning and coding performance.. Google released Gemini 3.8 Flash on Sept. 2, its fourth Flash model in under four months, targeting long-horizon software engineering, autonomous agents and complex reasoning. It supports a 1-million-token context window, 64K-token outputs and adjustable thinking levels. A restricted Flash Cyber variant detects vulnerabilities and automates patches; available through the Fairwind Program to vetted defenders, it reportedly scored 86.2% on CyberGym, underscoring the race for fast, affordable agentic AI—and the security risks of broader access. Story 3. Anthropic. Anthropic shipped Claude Fable 5.1 and Mythos 5.1, cutting cache-read pricing sharply and adding enterprise safety features for agentic and coding workloads.. Anthropic released Claude Fable 5.1 for general use and restricted Claude Mythos 5.1 to vetted cyber and life-sciences organizations. They are the same underlying model, differentiated by safeguards. Cache reads fell 75% to $0.25 per million tokens; pricing remains $10 input/$50 output. Both support 1M-token context, 128K outputs and adaptive reasoning. Story 4. Meta. Meta unveiled Muse Spark 1.3, a new agentic model aimed at coding and long-horizon computer-use tasks, with performance improvements over prior Muse versions.. Meta released Muse Spark 1.3 on September 2, positioning it as a stronger agentic model for coding and long-horizon computer tasks. Available in Muse Code and the Meta Model API, it supports a 1-million-token context window and text, image, and video inputs. Meta says it uses about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2, while scoring 75.4% on DeepSWE 1.1 and 88.8% on Terminal-Bench 2.1—raising competition in coding-agent efficiency and reliability. Story 5. Microsoft. Microsoft AI released MAI-Transcribe-2, a speech model aimed at voice-agent builders, claiming strong accuracy and high real-time throughput.. Microsoft AI released MAI-Transcribe-2 in public preview on Sept. 3, positioning it as a high-speed speech-to-text model for voice agents and other production workloads. It supports 60 languages, speaker diarization, word-level timestamps, language identification, keyword biasing and noisy audio. Microsoft reports a 5.2% average WER on FLEURS and pricing of $0.10 per audio hour. Independent Artificial Analysis measured 2.0% AA-WER and roughly 411× real-time throughput, intensifying competition in speech AI. That's all for today's AI news briefing.