AI News September 7, 2026: GPT-6 Astra, Fable 5.1, Gemini 3.8, and the outage that shut everything down
Alexandre
··
Reading time: 9 min
Click to enlarge
Credit: Reuters / Al Jazeera
Three frontier models launched in 72 hours by three different companies. Anthropic kicked things off on September 1 with Fable 5.1 and Mythos 5.1, Google followed on September 2 with Gemini 3.8 Flash, and OpenAI closed the week on September 3 with GPT-6 Astra, which Greg Brockman introduced as the beginning of "the AGI era."
On the numbers side, Anthropic signs a $35 billion cloud deal with Lambda, OpenAI prices Astra at $10 per million input tokens ($50 output), and Google holds Gemini 3.8 Flash at $0.75 per million input tokens through the end of 2026. All of this with the Trump administration officially backing OpenAI in the New York Times copyright lawsuit in the background.
Here's what happened.
Anthropic launches Fable 5.1 and Mythos 5.1, signs $35 billion deal with Lambda
Click to enlarge
Credit: Anthropic
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, two models built on the same architecture but with different safeguard levels. Fable 5.1 is the generally available flagship, while Mythos 5.1, with lighter guardrails, is restricted to vetted participants in Anthropic's trusted-access programs. The announcement came on the same day as a $35 billion cloud computing deal with Lambda, the Nvidia-backed cloud provider (Bloomberg, Quartz).
Fable 5.1 performance
The standout number: 52.6% on Terminal-Bench-Science 0.1, more than double Fable 5's result on the same benchmark (MarkTechPost). On code, Fable 5.1 gains 13% on Terminal-Bench 4.0 over its predecessor. The context window stays at 1 million tokens, max output at 128,000 tokens, and adaptive thinking (the model's extended reasoning mode) is always on.
Cache reads at -75%
The other announcement, less flashy but directly impactful for developers: cache read pricing drops from $1.00 to $0.25 per million tokens. In practice, that translates to roughly 25% lower cost on typical workloads, and up to 45% savings on complex agentic tasks that make heavy use of caching (The Decoder, VentureBeat).
Mythos: same model, fewer guardrails
Mythos 5.1 runs on the same engine as Fable 5.1, but with lighter safety constraints. Access is limited to participants in Anthropic's trust programs, initially cybersecurity defense researchers and life sciences researchers. Screening is individual and invite-only. The idea: let researchers explore the model's full capabilities without the restrictions that limit advanced use cases, while controlling who gets access (Anthropic).
$35 billion for Lambda
The Lambda deal covers approximately 350 megawatts of computing capacity at a new data center campus being built by Hut 8 in Nueces County, Texas. Nvidia leases the facility and supplies its next-generation chips. Lambda installs the hardware and resells the compute to Anthropic. The campus is expected to power up in Q1 2027. This deal adds to the $45 billion committed to Nscale in West Virginia and the $100 billion ten-year commitment to AWS (PYMNTS, SiliconAngle). The Claude ecosystem keeps expanding, from models to developer tools (full Claude Code overview).
While Anthropic lays its foundations, Google accelerates its release cadence at a dizzying pace.
Click to enlarge
Credit: Google
Google ships Gemini 3.8 Flash, third model in six weeks
Google unveiled Gemini 3.8 Flash on September 2, 2026, internal codename "Skimaki," positioning it as the company's "smartest workhorse" for coding, long-horizon reasoning, and agentic workflows. This is the third refresh of the Flash line in six weeks, a release pace that reflects Google's determination not to fall behind Anthropic and OpenAI on software development (Quartz, AI Business).
Benchmarks
On DeepSWE v1.1, Gemini 3.8 Flash hits 73.7%, up from 65.3% for version 3.7, an 8.4-point jump in three weeks. On Terminal-Bench 2.1, the score moves from 85.8% to 89.4%. The gains are concentrated in coding and multi-step reasoning tasks, exactly where developers need reliability (Incrypted, blog.google).
Flash Cyber: the AI that patches Chrome
Alongside the main model, Google launched Gemini 3.8 Flash Cyber, a specialized variant for vulnerability detection and automated patching. The Cyber variant patches 2.6x more Chrome vulnerabilities than its predecessor on Google's internal benchmarks (Tech Insider). Access is restricted to governments and critical infrastructure operators through the Fairwind program, a restricted application-based program similar in spirit to OpenAI's Daybreak.
Pricing: calm before the hike
Google is holding introductory pricing at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. After that, prices double. For developers building on Gemini, the signal is clear: lock in costs now or budget for 2027 (SammyGuru).
The model is available through the Gemini API, AI Studio, Android Studio, and Gemini Enterprise, as well as for Pro and Ultra subscribers. Alphabet's stock rose 0.7% in after-hours trading following the announcement (Shattered).
While Google iterates at full speed on Flash, OpenAI is preparing an announcement of an entirely different magnitude.
Click to enlarge
Credit: OpenAI / The New Stack
OpenAI launches GPT-6 Astra and declares the start of "the AGI era"
OpenAI launched GPT-6 Astra on September 3, 2026, calling it "the world's most intelligent and aligned model." OpenAI president Greg Brockman declared "Welcome to the AGI era" during the presentation, positioning Astra as a step toward artificial general intelligence. The model scores 97.6% on FrontierMath Tier 4 v2, 98.6% on ARC-AGI-3, and 100% on ExploitBench, the cybersecurity benchmark (CNBC, OpenAI).
The first "Critical" model
Astra is the first model to cross the "Critical" threshold of OpenAI's Preparedness Framework, the internal framework that assesses the risks of each model before production deployment. In plain terms: the model can generate functional cybersecurity exploits, and these capabilities are deliberately blocked in production. Advanced cybersecurity features are only accessible to selected enterprises through the Daybreak program (a restricted application-based access program for authorized organizations), comparable to Google's Fairwind or Anthropic's Mythos program (Fortune, CNET).
AGI or marketing?
100% on a cybersecurity benchmark does not mean artificial general intelligence. Several analysts point out that ExploitBench measures the model's ability to identify and reproduce known vulnerabilities, not to reason autonomously in novel situations (Yahoo/ATT). On broader benchmarks, Astra dominates Fable 5.1 and Gemini 3.8 Flash on most of OpenAI's internal evaluations, but each company tests on its own grids, making direct comparisons difficult.
Pricing and availability
Standard API pricing is $10 per million input tokens and $50 per million output tokens. A "Fast" mode at roughly double the standard rate promises 2.5x faster processing. The model is available to ChatGPT Plus ($20/month), Pro ($100-$200/month), Business and Enterprise subscribers, as well as through the API, Microsoft Azure, Amazon Bedrock, and AWS. Advanced cybersecurity remains restricted to the Daybreak program. Computer use (the model's ability to control a computer end-to-end) is one of the highlighted capabilities for research, coding, and science tasks (PCMag, The New Stack).
For developers working with multiple providers, the week is dizzying: three frontier models in 72 hours, each with its own specializations and restricted access programs. The detailed comparison between these architectures and how they work in a coding workflow (AI code agent architecture) only confirms that multi-model diversification has become an operational necessity.
ChatGPT, Claude, Grok, and Gemini go down at the same time for 90 minutes
On September 3, 2026, just hours after GPT-6 Astra's launch, the four major AI chatbots experienced a simultaneous outage. ChatGPT, Claude, Grok, and Gemini all went offline within the same 90-minute window, leaving millions of users without access. Downdetector registered a massive spike in reports, and Microsoft Azure confirmed an incident in its East US region (Quartz, Economic Times).
Azure: the common thread
Post-incident analysis points to a regional outage in Azure East US, the Microsoft cloud region that hosts ChatGPT, Claude, and Grok. All three services depend on Azure infrastructure for all or part of their operations. Gemini, which runs on Google's own infrastructure, recovered faster than the others. Cloudflare and AWS also reported related disruptions within the same time window (Shattered, Tech Insider).
Timeline
The outage started in the early morning hours (Pacific time). OpenAI confirmed elevated error rates for ChatGPT and Codex. Anthropic reported outages across multiple Claude models, including Opus 4.8 and Opus 5. xAI acknowledged that Grok was down. First signs of recovery appeared around 8:49 AM PT, with full restoration of ChatGPT, Claude, and Grok confirmed by 12:38 PM PT. Gemini returned to normal shortly after.
What this reveals
Four competitors, one shared point of failure. The AI industry's heavy reliance on Azure East US creates a systemic risk that few users appreciate. When three of the four leading chatbots share the same critical infrastructure, a regional Azure outage does not hit one service, it paralyzes the ecosystem. The incident falling on the very day of Astra's launch adds a layer of irony to OpenAI's "AGI era" messaging: the world's most advanced artificial intelligence remains dependent on a data center in the eastern United States.
The incident is a reminder of why infrastructure diversification remains just as critical as model diversification (AI news August 29).
Click to enlarge
Credit: Bloomberg Law
Trump administration backs OpenAI against the New York Times on copyright
On September 2, 2026, the U.S. Department of Justice filed a 20-page brief in support of OpenAI in the copyright lawsuit brought by the New York Times. This is the first time the federal government has officially intervened on behalf of an AI company in an intellectual property dispute. The DOJ characterizes training models on copyrighted content as "extraordinarily transformative" and invokes fair use, the legal doctrine that permits the use of protected works under certain conditions (Bloomberg Law, Politico).
The "national security" argument
The DOJ brief goes beyond simple fair use. It argues that a New York Times victory could "threaten national security" by slowing U.S. AI development in the face of Chinese competition. The department also argues that a licensing requirement would disproportionately benefit large publishers at the expense of small newsrooms, which would lack the resources to negotiate deals with AI companies (Nieman Lab, Quartz).
Implications beyond the press
This stance extends well beyond the New York Times case. The music industry, which is fighting its own legal battles against Anthropic and Suno, is watching closely. If courts accept fair use for model training on news articles, the precedent could apply to music, images, and all creative content used as training data (Music Business Worldwide).
The timing of this intervention is notable. Sam Altman spoke at the G20 Innovation Ministerial in Chapel Hill on September 2, the same day the brief was filed, where he declared AI "non-negotiable" for economic growth (News Observer). The proximity between OpenAI and the U.S. administration is written on every page of the brief.
Other news in brief
Sam Altman at the G20: "AI is non-negotiable": at the G20 Innovation Ministerial in Chapel Hill on September 2, OpenAI's CEO said future AI systems will act as "persistent virtual collaborators." He also stated that "the work that used to take a startup three months is now doable in 17 minutes with Codex" (ETV Bharat).
Lyte, founded by ex-Apple Face ID engineers, raises $165 million: the AI and robotics startup tripled its valuation to $1.6 billion. The funding round confirms continued investor appetite for AI applied to vision and robotics (Bloomberg).
Crusoe hits $30 billion valuation: the AI infrastructure cloud startup closed a new funding round on September 4, bringing its valuation to approximately $30 billion.
Texas halts AI data center power contracts: the Texas public utility commission announced a moratorium on new power contracts for AI data centers, citing "ghost demand" (projects that reserve energy capacity without ever getting built). The signal is clear: AI infrastructure growth is hitting the limits of the power grid (The Star).
Conclusion: the week everything accelerated
Three frontier models in 72 hours, $35 billion in cloud contracts, the U.S. government taking a position on AI copyright, and an outage revealing that the entire ecosystem runs on a handful of Azure data centers. The pace has become so relentless that each announcement gets less than 24 hours of exposure before being eclipsed by the next one.
The real question this week is not about benchmarks. It is about infrastructure. When three of the four leading AI services share the same point of failure, model progress becomes secondary to the resilience of what makes them run. The $180 billion Anthropic has committed to infrastructure deals in 2026 (Lambda, Nscale, AWS) shows the company gets it.
Did you feel the September 3 outage? Are you diversifying your AI tools or sticking with a single provider? Send me a message on Twitter/X or drop a comment.
Alex
Key takeaways
Fable 5.1 and Mythos 5.1 ship Sept 1, cache reads down 75%, 52.6% on Terminal-Bench-Science
Gemini 3.8 Flash is Google's 3rd model in 6 weeks, pricing held at $0.75/M tokens through end of 2026
GPT-6 Astra scores 97.6% on FrontierMath and 100% on ExploitBench, priced at $10/M input tokens
ChatGPT, Claude, Grok and Gemini go down simultaneously for 90 min, Azure East US to blame
Trump administration files brief backing OpenAI in NYT copyright lawsuit
