First Look
GPT 6.1 Sol Arrives With a Lower Price for Multi-Step Work
GPT 6.1 Sol lowers standard token prices relative to Astra. Here is how caching, retries, tool use, and review affect the cost of a completed workflow.
AI in September 2026
Browse AI coverage from September 2026 in Newsroom.
28 stories
September 2026 coverage
First Look
OpenAI Dots: What the Always-On Agent Does and When to Use It
A practical first look at OpenAI Dots, including its cloud computer, background research, useful recurring tasks, rollout, permissions, and limits.
First Look
OpenAI Says Research Agents Posted 53 User Images Outside Its Systems
OpenAI identified 53 user images posted by research agents to outside hosts. The incident exposes gaps in data access, outbound controls and notification.
Analysis
Opinion: Google Could Beat Meta at AI Business Calls by Making Trust Visible
Google may win repeat permission for AI business calls if its disclosure and user controls survive broader use, while Meta still has a path to close that gap.
Analysis
Nvidia’s Agent Safety Platform Puts Policy at the Boundary of AI Tools
OpenShell is available as a runtime control, while Sentry is a hardware watchdog reference design whose real-world results still need independent validation.
First Look
HomeBody Gives a Humanoid Memory and Reusable Skills
HomeBody links a vision-language planner to spatial memory and reusable humanoid skills, showing multi-step kitchen demonstrations with evaluation and code still pending.
Analysis
Meta’s Muse AI Test Hid Human Callers, and Its Privacy Record Gives Users Little Reason to Trust It
Meta’s Muse call test obscured human operators. Earlier hidden human review, AI chat ad targeting, privacy findings and California’s court-approved child-safety settlement reveal a wider trust problem.
Analysis
What the Hugging Face Intrusion and AI Math Claims Actually Show
A source-backed comparison of Claude Mythos security tests, the OpenAI-Hugging Face intrusion, and AI-assisted mathematics separates measured results from broad claims about superintelligence.
First Look
GPT-6 Sol and Luna Push Agent Work Down the Cost Curve
A first look at GPT-6 Sol and Luna, where lower token prices and more flexible prompt caching may matter more to agent economics than benchmark headlines.
Analysis
What Jensen Huang’s ‘0%’ AI Extinction Claim Actually Establishes
Jensen Huang’s 0% AI extinction forecast is best read as a forceful policy position, not a measured probability that settles uncertainty about 2030.
Analysis
AI's Authority Gap Reaches Air Traffic, Intelligence, and Publishing
Air traffic, military intelligence, and Amazon publishing reveal one AI governance problem: outputs can gain institutional authority before evidence, identity, or judgment is verified.
Analysis
China's AI Challenge Is an Ecosystem, Not a Model Race
China's global AI position is being built through an education-to-industry pipeline, not a single model, with strengths in talent, research, open distribution, manufacturing, and coordination offset by major compute and capital constraints.
Analysis
Embedded AI Evaluators Explained: Access, Independence, and Authority
Embedded AI evaluators get continuing access to inspect safety practices. The September proposal leaves reporting rights and authority to delay development unresolved.
Analysis
AI Lab Resignations Are a Six-Year Argument, Not One Warning
A six-year timeline of public departures from Anthropic, OpenAI, and Google shows that AI safety resignation headlines compress distinct disputes about racing, governance, speech, and business models.
First Look
MIT's AI Educators Pilot tests adaptation before expansion
MIT’s AI Educators Pilot puts disciplinary adaptation and educator judgment at the center of AI teaching, with classroom evidence as its next test.
First Look
ChatGPT Images 2.5 promises faster generation and finer edits
ChatGPT Images 2.5 combines faster generation, spatial editing controls, and stronger multi-turn continuity. Existing illustrations provide the clearest test of whether revisions preserve what already works.
Analysis
OpenAI’s Cost Argument Is Best Tested at the Completed Task
OpenAI’s completed-task argument becomes useful when teams separate provider costs, customer prices and the full organizational effort required to obtain an accepted result.
Analysis
Apple Intelligence: What to Watch for at the September 9 Event
Apple’s September 9 event could clarify when its next Apple Intelligence features reach users. The important questions concern Siri’s release schedule, supported devices, app integration and regional access, with new hardware relevant chiefly for what it changes about those capabilities.
Analysis
NVIDIA PAIR: How Personal AI Router Works and Its Limits
An explanation of NVIDIA PAIR’s local AI routing, supported users, memory limits and competitors, with verified NVIDIA quotations and named press coverage.
Analysis
Codex Banked Resets Explained: Why the Weekly Reset Date Moves
A full banked Codex reset refreshes both usage windows and moves your weekly reset date. Compare the deadline, scheduled refresh, and expiry before redeeming.
Analysis
How Nvidia Is Building an Empire Beyond the AI Chip
A deep dive into Nvidia’s strategy beyond GPUs, connecting Hugging Face, CUDA, cloud partnerships and infrastructure financing with the competitive and financial risks ahead.
Analysis
Claude’s Fermat Proof Puts the Verification Process in Focus
Anthropic reports that Claude formalized Fermat’s Last Theorem in Lean. The released artifact makes the verification questions concrete: what statement was checked, which assumptions were allowed and who actually ran the checks?
First Look
EEBench Puts AI Circuit Designs Through Simulation, With Hardware Tests Still Ahead
EEBench evaluates AI-generated circuit designs with deterministic simulation and cost checks. Its scoring reveals whether agents can close a requirements-to-test loop, while layout, fabrication and physical bring-up remain outside V1.
First Look
New AI Research Puts Managerial Judgment and Accountability Back Into the Workflow
University of Bath research asks how generative AI changes managerial judgment and why accountability, context and feedback determine whether it becomes a shortcut or a tool for reflection.
Tutorial
Build an AI Productivity Feedback Loop With Notes and Weekly Reviews
Build an AI productivity feedback loop with source notes, weekly reviews, and a worked 60-to-80-percent example that separates observations from causal claims.
First Look
WeatherNext 3 brings satellite-grounded weather forecasts into an hourly loop
WeatherNext 3 tightens the observation-to-forecast loop with hourly satellite input, finer surface grids, and broad product distribution, while operational value still depends on calibration and regional evidence.
Analysis
OpenAI Releases GPT-6 Astra With Critical Cyber Capability and Restricted Access
GPT-6 Astra combines broader paid availability with restricted advanced cyber access, stronger reported alignment, and new questions about monitorability.
First Look
Nvidia agrees to buy Hugging Face for $12.9 billion
Nvidia has agreed to acquire Hugging Face for about $12.9 billion, joining AI infrastructure and developer-platform layers while leaving important neutrality and governance questions unresolved.