OpenAI rolls out GPT-6 Astra to trusted users
OpenAI has begun limited rollout of GPT-6 Astra, its new flagship model with 1.05 million token context, self-reported benchmarks including GPQA 96.0, and a $10/$50 per million token pricing tier. The model is positioned for high-stakes agentic and long-context work but starts with restricted trusted-access only. Primary source: LLM Stats AI News.
What else matters
Anthropic releases Claude Fable 5.1
Anthropic launched Claude Fable 5.1 on September 1, adding it to major gateways. The update continues the rapid cadence of frontier model iterations across providers. Source: LLM Gateway Timeline.
Seven new models tracked this week
Release trackers logged GPT-6 Astra (OpenAI), Muse Spark 1.3 (Meta), multiple Gemini 3.8 variants (Google), Qwen3.8-Max, and Claude Fable 5.1. The pace shows no sign of slowing. Source: AI Release Tracker.
Benchmark contamination remains a live concern
New arXiv papers highlight ongoing issues with test-set leakage into training data and the limitations of linear probes for detecting concepts in LLMs. These papers underscore persistent evaluation challenges. Sources: multiple arXiv preprints via LLM Stats.
Open-source momentum in enterprise
Hugging Face reports roughly half of Fortune 500 companies now prefer self-hosted open models over rented APIs for cost, privacy, and customization. This shift is material for deployment strategy. (Continuing coverage.) Source: AI Herald.
*Worth your time:* The full set of new arXiv papers on agent scaffolding, personalized agents, self-repair, and RAG improvements (linked in LLM Stats).
