Update: PyTorch Foundation added an Introduction Track and a full-day PyTorch Associate Training (Oct 19, San Jose) to PyTorch Conference North America 2026.
Guide refreshed for late September: OpenAI's GPT-6 Sol and Luna release and Anthropic's Claude Opus 5.5 launch (22 September) added to the landed section, with the lead frame, watch thread on API pricing, unknown list and sources updated.
Baseline table refreshed to 22 September 2026 with GPT-6 Sol / Luna and Claude Opus 5.5 rows; the earlier GPT-5.6 Sol and Opus 5 rows are kept for comparison, and the volume sketch, analysis and sources were updated to match.
Availability update: Qwen3.8-Omni-Flash is live on Qianwen / Model Studio as of September 18 — title, lead, and Confirmed availability refreshed from upcoming to launched.
Update: Project code is on GitHub and a live Paper2Agent demo is at paper2agent.ai/live, with hosted AlphaGenome/Scanpy/TISSUE MCP endpoints linked from the repo.
Update: OpenAI's CFO indicated in a Sept 14 Fortune follow-up that the IPO is now expected in 2027 or later, beyond Altman's earlier 'not 2026' statement.
Update: IBM's Hugging Face blog states that as of September 8, 2026, PatchTST-FM-r2 is the top-performing zero-shot time-series model released under a permissive, commercial-friendly open-source license.
Added AI Infra Summit claims under Confirmed/Unknown: Groq 3 LPX and SemiAnalysis AgentX throughput-per-megawatt figures (NVIDIA-attributed), plus a link to the summit post.
We did not copy the twin draft’s long partner and CPU benchmark list into this article.
Added Monday’s verified cybersecurity rally (CrowdStrike ~14%, Palo Alto Networks ~13%) and CrowdStrike CEO George Kurtz’s response to the pace debate; sharpened the All-In Trump–Huang stage exchange from additional reporting.
Official DeepSeek news page (10 Sep) confirms V4.1-Flash is live as deepseek-flash with native multimodal support; discloses 552B MoE Causal Encoder–Decoder (8B/16B active), KV-cache footprint cuts, retirement of V4-Flash/Vision-Exp, and Pro→Flash remapping from 04:00 UTC on 14 Sep at Flash rates until V4.1-Pro.
Body, Confirmed/Unknown, and Sources refreshed from that primary.
Meta officially announced and launched Muse on September 8, 2026, confirming the personal AI agent is available in the US via a dedicated app and WhatsApp with free and paid subscription tiers.
The closed alpha described in the article has progressed to a public launch.
Updated Sep 8, 2026: A recent Atlantic Council analysis found 75 major data-center projects worth over $130 billion were delayed or canceled in the first three months of 2026 due to local opposition, underscoring the domestic backlash's real-world impact.
OpenAI said on 2 September 2026 that Astra was not part of the July Hugging Face breach but that the incident prompted stronger pre-release safeguards and reiterated limited early access before Daybreak Blue rollout.
31 Aug 2026: DeepSeek open-sourced V4-Flash-Vision-Exp on Hugging Face under MIT license (weights, tokenizer, and minimal PyTorch inference reference).