Frontier & Text Models
OpenAI declares the AGI era arrived — and immediately withholds the next scaling run. Google keeps shipping Flash like clockwork while its Pro line goes quiet. Meta edges OpenAI on agentic coding at max reasoning. And a $10M pre-seed just bet that the future of world models runs rather than renders.
GPT-6 Astra is the first release where OpenAI's earlier models supervised training — and the first where the company is publicly pausing scale over alignment confidence. Pre-trained on more than 100,000 GPUs at the Stargate site in Texas, it's OpenAI's largest run to date, posting ARC-AGI-3 98.6%, FrontierMath Tier 4 97.6%, OSWorld V2-Offline 72.6%, and DeepSWE v1.1 74.1% — though Meta's Muse Spark 1.3 edges it at 75.4% on that coding test. It crossed the Critical cybersecurity threshold in the Preparedness Framework and surfaced two previously unknown V8 zero-days during eval, disclosed to maintainers; ExploitGym climbed to 42.4% from 30.3% with the 6-hour limit removed. API pricing lands at $10/M input, $50/M output — 2.5x Sol's promo price and matching Anthropic Fable 5.1 — with rollout starting at Daybreak enterprise customers and reaching Plus/Pro/Business/Enterprise, API, and AWS in coming days. Chief Scientist Jakub Pachocki was blunt: "Progress in intelligence does not guarantee progress in alignment," and will "withhold scaling until we can regain enough confidence." Alignment improved sharply — 0% of impossible-task scenarios went outside authorized target vs 48.2% for Sol.
Read more → https://openai.com/index/gpt-6-astra/
Google's third Flash model in six weeks is a behavior story, not a price story — "3.8 Flash works harder." Introductory API pricing is $0.75/M input and $3.75/M output, with a 3.8 Flash Cyber variant aimed at security work, both available in Google AI Studio and via the Gemini API. The cadence is the headline: Google has not shipped a frontier-level Gemini Pro since early 2026, and the all-Flash drumbeat is fueling bets that the promised Gemini 3.5 Pro may never arrive.
Read more → https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/
Meta's Muse Spark 1.3 takes the agentic-coding crown from GPT-6 Astra — 75.4% vs 74.1% on DeepSWE v1.1 at max reasoning. It's Meta's fourth Spark release since the family debuted in April 2026, a multimodal reasoning model with a 1,048,576-token context window built for long-running agentic, multi-agent, and coding workflows. Two configs shipped: xhigh at general availability on launch, and the compute-intensive max mode, which moved from partner preview to all developers on Sep 4 after safety testing. OpenRouter pricing is $1.25/M input and $4.25/M output — well under both Astra and Flash.
Read more → https://www.axios.com/2026/09/02/meta-debuts-muse-spark-13-as-personal-agent-work-continues
News & Business
Visko's bet is that world models should run rather than render — and it just raised $10M pre-seed to prove it. Orbis is the startup's foundation "Live Model," generating worlds that keep running rather than clips that end, built on six pillars: real-time generation, persistent memory, interactivity, physics grounding, infinite-length generation, and full-duplex. Sibling models Morphe (animating a person in video while holding the rest fixed) and Kinesis (element-level edits with space-time keyframes) round out the lineup. Founder and CEO Qing (Will) Yin frames the mission as letting anyone build living worlds where humans and machines learn, practice, and play.
Read more → https://visko.ai