OpenAI and Anthropic both restricted their strongest models on September 1, 2026 over offensive cyber capability. Two days earlier the War Department added ChatGPT Mil and Grok for Government to GenAI.mil for roughly 3 million1 users.
On August 31, 2026 the War Department added OpenAI's ChatGPT Mil and Starshield AI's Grok for Government to GenAI.mil, joining Google's Gemini in an IL5 environment for roughly 3 million1 military and civilian users. Procurement moved while the labs were tightening access.
Meta calls Muse Spark 1.3 its most capable model to date but has not decided whether to publish the weights, even as an open release for 1.2 stays on the calendar. Coverage citing Bloomberg puts 1.3 near Claude Fable 5.1, with the EU AI Act's Article 53 exemption entangled in the call.
Cerebras reported roughly 750 tokens per second on GPT-5.6 using 900,0002 cores and 44GB of on-chip SRAM, claiming 21 petabytes per second of memory bandwidth with no quantization. Inference latency is becoming a hardware architecture question, not a serving trick.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →OpenAI and Anthropic each restricted their strongest systems on September 1, 2026, citing the same reason: the models are too good at offensive cyber work. OpenAI says Astra is the first model to cross the 'Critical' cybersecurity threshold under its Preparedness Framework, able to find unknown flaws and build...
Continue readingCENTCOM said U.S. forces completed a wave of strikes on Islamic Revolutionary Guard Corps targets across Iran, widening a campaign previously centered on launcher sites near the Strait of Hormuz.
Mood reads Greed, with capability claims from Cerebras at roughly 750 tokens per second on GPT-5.6 outrunning the legal overhang from Sony Music and Warner Music's new suit against Anthropic.
The same week two labs judged their best models too dangerous to release freely, the War Department opened ChatGPT Mil and Grok for Government to roughly 3 million1 personnel: the gating is happening at the lab, not the customer.
GenAI.mil now offers ChatGPT Mil, Grok for Government, and Google's Gemini to roughly 3 million1 military and civilian users in an IL5 environment. The government is standardizing on multiple commercial vendors rather than a single bespoke stack, covering planning, policy drafting, and logistics. That makes vendor lock-in less likely and lab-side safety policy more consequential for a captive user base.
Meta has not decided whether to publish Muse Spark 1.3's weights, though it still plans an open release for version 1.2. Coverage citing Bloomberg calls 1.3 competitive with Claude Fable 5.1 and stronger than GPT-5.6 Sol at coding. Every increment in capability raises the cost of the open-release commitment, which is exactly the calculus OpenAI and Anthropic just resolved in the other direction.
750 Tokens Per Second, Same Weights
Read full analysis
Deep Dive
7.1/ 10
One payment from 7.99 euros turns the street you are standing on into a murder mystery, and the only real question is whether the clues match the pavement.
The mechanic is simple and it mostly works: pay once for a single quest, walk between real landmarks, solve a puzzle at each stop for 1.5 to 3 hours, no guide and no reservation. That makes it a genuinely good default for a spare afternoon in a new city or a date night you do not want to schedule, especially since the quest is yours forever and runs offline. The catch is quality variance across 3,000+ quests written by 30,000+ creators, and the loudest public complaints are exactly about that: clues that feel bolted onto locations rather than drawn from them.
“A TripAdvisor reviewer called a trivia-style outing quite disappointing, saying there were no real clues actually tied to the places you visit and it felt like just an app.”TripAdvisor
Wialon Platform added a ChatGPT extension that lets partners query their own account, vehicle, and operations data in natural language instead of building a report. It is an operations assistant rather than a coding tool, so the honest test is a question you currently answer by exporting to a spreadsheet: ask it once, then verify the number against the source table before you trust it in a customer conversation.
Try itOrange AI decodes and summarises video comment threads, which is the kind of feedback pile most creators abandon after the first hundred replies. Note that the directory listing also advertises agent building and team task management, so the scope is inconsistent: open the entry, confirm which product you are actually signing up for, and test it on one video's comments before wiring it into anything.
Try itThe 2026 wave 1 plan page now carries a notice that release plans stop being published in September 2026, with new Dynamics 365, Power Platform, and Dataverse capabilities moving to the AI at Work roadmap. If your change management process depends on scraping or reading those wave plans, that pipeline breaks this month, so repoint your internal watch list before a feature ships you did not see coming.
Try itThe YC industry page currently shows 909 AI-funded startups out of a broader directory of more than 5,00014 companies, with batch year, headcount, and status on each card. Use the employee count and Active flag as a first cut: a two-person S2026 company and a 500-person S2016 company like Scale AI carry very different procurement risk, and the page tells you which is which before you book a demo.
Try itWrite a capability downgrade plan. List every feature in your stack that depends on a frontier tier, then name the fallback model for each one. The September 1 restrictions on Astra and Claude Mythos 5.1 show access tiers can narrow with no notice.
Ship a plain-language query layer. Copy the Wialon ChatGPT extension pattern: expose one operational dataset behind natural language questions instead of a dashboard, and cap it to read-only queries first. Orange AI's YouTube comment reader is the same move on unstructured input.
Audit your Power Platform release dependencies. Microsoft stops publishing Dynamics 365 and Power Platform release plans in September 2026, so pull the 2026 wave 1 page now and archive anything your roadmap cites. Replace plan links with your own tracked changelog.
Two frontier labs restricted their best systems on the same day for the same reason, which makes offensive security the first capability to actually cost users access. Meanwhile the War Department wired three commercial models into an IL5 environment for 3 million people, and Sony and Warner Music sued Anthropic in the same week it shipped Fable 5.1. Builders should assume the top tier is now a privilege with conditions attached, not a product line. Design around the model you are guaranteed, then treat frontier access as an upgrade you can lose.
One email per day. Unsubscribe anytime.