Google Gemini Gems: 90% of Users Are Missing the Best Features
September 13, 2026
A researcher who spent three years training models at both OpenAI and Anthropic resigned
September 13, 2026
Link to our newsletter: https://bitbiased.ai/
Three of the biggest AI systems went down on the same morning. Anthropic’s Claude started throwing elevated errors, xAI’s Grok went completely offline minutes later, and OpenAI’s ChatGPT and Codex followed. For over an hour, the incidents overlapped — immediately triggering theories about a coordinated cyberattack or a massive shared cloud failure.
But that’s not what the evidence shows.
OpenAI blamed a routing error. Anthropic cited an infrastructure issue affecting Claude.ai, Claude Code, Claude Cowork, and its API. SpaceX tied Grok’s outage to its Memphis compute center. Meanwhile, AWS, Microsoft Azure, Google Cloud, and Cloudflare showed normal operations. Even reports that Google Gemini had gone down were never confirmed by Google.
Three companies. Three different explanations. One very strange morning.
There is still an interesting complication. SpaceX operates compute infrastructure reportedly used by other AI companies, including Anthropic. That raises the possibility that apparently separate incidents could still expose how concentrated modern AI infrastructure has become. Nobody has confirmed that connection, and the companies haven’t released detailed postmortems, so major questions remain.
But the outages were only the beginning of a much bigger week.
On September 1, OpenAI revealed that its upcoming model, internally called Astra and expected to ship as GPT-6, had crossed into the “Critical” cybersecurity capability tier under its Preparedness Framework. In testing, Astra reportedly discovered two real, previously undocumented zero-day vulnerabilities autonomously.
OpenAI isn’t shelving the model. Instead, it plans to restrict initial access through a vetted group called Daybreak Blue, giving trusted testers and security partners access before any broader rollout.
Anthropic has made a strikingly similar decision.
Claude Fable 5.1 is publicly available and can identify software vulnerabilities, but according to Anthropic it cannot exploit them. The more capable Claude Mythos 5.1 is restricted to approved researchers working in cybersecurity and life sciences.
OpenAI and Anthropic use different safety frameworks, but both appear to be reaching the same conclusion: the most cyber-capable frontier AI models may no longer go directly from the lab to unrestricted public access.
Then there’s the Pentagon.
On August 31, the Department of Defense expanded GenAI.mil, its secure AI platform built for roughly three million DoD personnel. Google’s Gemini for Government was already integrated. Now ChatGPT Mil and Starshield AI’s Grok for Government are joining it.
The new systems are IL5 and CUI-certified and designed for controlled unclassified government work, including document drafting, planning, logistics, file handling, and acquisition support. More than 1.7 million people had reportedly enrolled on GenAI.mil by summer 2026, before the OpenAI and xAI additions went live.
OpenAI also signed a roughly $200 million Pentagon contract in June 2025 to develop government-grade AI tools.
But one major frontier lab is missing: Anthropic.
Claude’s absence reportedly follows the Trump administration designating Anthropic a “supply-chain risk” after a separate policy dispute, a designation Anthropic is challenging legally. That means access to one of the largest government AI deployments is increasingly being shaped not only by benchmarks and technical capabilities, but by policy.
Put these stories together and a bigger pattern emerges.
AI is becoming operational infrastructure while the systems themselves are becoming powerful enough that their creators are restricting access. At the same time, governments are deciding which frontier labs get plugged directly into enormous institutional deployments.
The conspiracy surrounding the outages wasn’t supported by the evidence. What actually happened may be more important: three major AI providers failed at nearly the same time for apparently different reasons, two competing labs independently decided their newest models needed restricted access, and the Pentagon picked its AI partners — while leaving one of the biggest names in AI outside the door.
CHAPTERS
00:00 Three AI Giants Went Down at the Same Time
01:55 Three Companies, Three Different Explanations
04:34 The Conspiracy That Wasn’t
07:47 Why Two Labs Just Locked Down
09:15 Anthropic’s Quieter Version of the Same Decision
11:21 The Pentagon Just Chose Its AI Partners
14:06 The Big Picture
#openai #anthropic #chatgpt #claude #ai
