Summary
AI Search delivers a comprehensive weekly roundup covering one of the most active model-release weeks on record. The headline is OpenAI’s GPT-6 Astra, described as significantly outpacing the competition, but the week also saw Anthropic’s Claude Fable 5.1 briefly hold the top spot before being overtaken, Google’s Gemini 3.8 Flash, Meta’s Muse Spark 1.3, and new releases from Deepseek and Alibaba’s Qwen line — all shipping within days of each other.
Gemini 3.8 Flash receives the most detailed benchmark treatment: it ranks first on the DeepSWE long-horizon agentic software engineering leaderboard, achieves inference speeds around 348 tokens per second (far above competitors), and scores state-of-the-art on multimodal benchmarks for scientific figure understanding, long video comprehension, and biology research tasks. The video flags a key caveat — Gemini 3.8 Flash uses significantly more output tokens than rivals, which complicates its cost-efficiency claims. It ranks eighth on the Arena leaderboard and outside the top ten on LiveBench.
The open-source coverage is equally detailed. H3 World converts the Minimax H3 video model into an interactive game engine by translating keypresses into natural language descriptions fed through the model’s existing language layer. Solar WM extends the same concept across multiple video generators including LTX 2.5 and Minimax H3, sustaining coherent interactive worlds for over an hour, and releases its full training dataset of 1.3 million clips (25TB) publicly. Google’s TimesFM3 — a 330 million parameter zero-shot time-series forecasting model pre-trained on over one trillion data points — rounds out the open-source segment, ranking first among time-series foundation models on GiftEval and other standard benchmarks.
📺 Source: AI Search · Published September 06, 2026
🏷️ Format: Roundup
