GPT-6 Astra is out. Critical cyber still stands. Tip shorthand “Astra 6” is not the official name. OpenAI’s pages call it GPT-6 Astra (API gpt-6-astra). On 3 Sep it started rolling to a limited set of orgs, then Plus/Pro/Business/Enterprise, API, Azure, and Bedrock. Help notes that day still said not generally available yet. The 1 Sep Path to Astra post and the release System Card both still say it meets the Critical cybersecurity tier under the Preparedness Framework. First OpenAI model at that level. Advanced cyber stays gated (testers, then Daybreak Blue). Path to Astra also says the Critical evals shown used Daybreak Blue access, not default production. Source: https://openai.com/index/gpt-6-astra/ Also: https://openai.com/index/path-to-astra/ System Card: https://deploymentsafety.openai.com/gpt-6-astra
Grok Bot is a cloud teammate, not Grok chat xAI’s Grok Bot is a separate product from grok.com chat: named Bots on a per-user persistent cloud computer with browser, filesystem, and terminal. Connectors/MCP where they exist; computer use where they do not. Official ship window: launch 11 Aug, more plans 26 Aug, X connector 29 Aug, Enterprise + design essay 3 Sep. List prices on x.ai/bot: Cursor Individual from $20/mo, SuperGrok from $30/mo, Teams $40/seat/mo. Hard caveat: every Bot on one account shares one computer — that is not a security boundary between Bots. Source: https://x.ai/bot Also: https://docs.x.ai/grok-bot/overview Design essay: https://x.ai/news/designing-grok-bot
Anthropic ships Fable 5.1 wide, Mythos 5.1 behind a gate Anthropic says Fable 5.1 and Mythos 5.1 are the same model with different safeguards. Fable 5.1 is out for general use. Mythos 5.1 is only for vetted US groups in cyber and life sciences programs. List price stays $10 per million input tokens and $50 per million output. Cache reads dropped to $0.25, which Anthropic says cuts typical bills about 25%. Fable 5.1 may find software bugs. It is still blocked from writing exploits. Source: https://www.anthropic.com/claude-fable-and-mythos-5-1
Google’s 3.8 Flash is cheap. The cyber cut is not public. On 2 Sep 2026 Google posted Gemini 3.8 Flash at the same intro price as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output, through 31 Dec 2026. A second cut, 3.8 Flash Cyber, is only for “trusted defenders” in a new Fairwind Program. Google says Chrome Security got 2.6 times more correct patches from it than from much larger commercial models. That is Google’s claim, not an outside audit. Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/
MISO asked FERC to write large-load rules into the tariff On 28 Aug 2026 the Midcontinent Independent System Operator filed “Interconnection Reliability Requirements for Large Loads” at FERC, docket ER26-3650. The ask is for the rules to take effect 4 Dec 2026. Driver: data centers, AI compute, and other big loads. The filing exists. Coverage puts a large-load line around 50 MW; that number was not read from the FERC PDF this run. Source: https://elibrary.ferc.gov/eLibrary/filelist?accession_number=20260828-5278&optimized=false Also: https://www.misoenergy.org/planning/large-loads---container-page/large-load-additions/