GPT-6 Astra, iPhone Duo, $6B AI Funding: Tech's Wildest Week

Some weeks the technology industry drips news. In the first ten days of September 2026 it opened a fire hose: a frontier model that its own maker linked to AGI, the first complete brain map of a male animal, Apple's first folding iPhone, $6.13 billion of AI funding in 48 hours, and, running underneath all of it, the loudest safety warnings the field has produced. Here is the whole week in one place, with the numbers that matter.

The AGI week: GPT-6 Astra lands, and Jensen Huang calls it

OpenAI released GPT-6 Astra on September 3, first to a limited "Daybreak Access" group and then broadly the following day. President Greg Brockman called it a "generational leap" and said it may eventually be seen as the arrival of artificial general intelligence. The specs back the swagger: a 1,050,000-token context window, near-perfect scores on key reasoning benchmarks, and API pricing of $10 per million input tokens and $50 per million output.

Three days later, Nvidia CEO Jensen Huang turned the release into a declaration. "GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived," he posted on X on September 6. The pushback was immediate. Gary Marcus called the claim out as arriving "with no evidence and no definitions," and the exchange restarted a definitional argument the industry has never actually settled.

Then, on September 10, DeepSeek reminded everyone that frontier capability no longer stays expensive for long. Its new open-weight V4.1-Flash is a 552-billion-parameter mixture-of-experts model that keeps only 8 billion parameters active while reading a prompt, and it scored 90.6 on Terminal-Bench 2.1, narrowly ahead of Claude Opus 5 (89.1) and GPT-5.6 Sol (88.8). Off-peak, a cached input token costs $0.003 per million. That is not a typo, and it is aimed straight at the economics of the closed labs.

$6.13 billion in 48 hours: the money kept pace

The venture market answered the model news with the densest two days of AI fundraising this year. Between September 8 and 9, Cognition, Mistral AI and Harvey together raised $6.13 billion, at a combined post-money valuation of $87.5 billion. (Method: the sum of the three announced rounds, with Mistral's €3 billion converted at the rate implied by its reported $24 billion valuation.)

Cognition raised $2 billion at $48 billion, nearly doubling its May valuation of $26 billion while run-rate revenue climbed from $492 million toward $900 million. Mistral closed €3 billion led by Samsung, which the company describes as the largest equity round ever completed by a European technology company. And Harvey took $550 million at $15.5 billion, less than six months after being valued at $11 billion, with the explicit goal of training its own legal models instead of renting intelligence from OpenAI and Anthropic.

The pattern across all three: investors are no longer pricing a winner-take-all outcome. A coding agent, a European sovereign-AI champion and a legal vertical each got mega-round treatment in the same week that the general-purpose frontier moved.

The week, day by day

Date	What happened	The number that matters
Sept 1	World Labs launches Atlas, an omnimodal world model	$1.2B raised to date
Sept 3	Google and HHMI Janelia publish the complete male fruit fly connectome in Cell	166,000 neurons, ~125M synapses
Sept 3	OpenAI releases GPT-6 Astra	1.05M-token context
Sept 6	Jensen Huang declares "AGI has arrived"	trained on ~100K Grace Blackwell GPUs
Sept 8	Cognition raises Series E	$2B at $48B
Sept 8	Mistral AI closes Samsung-led round	€3B at ~$24B
Sept 9	Apple unveils the iPhone Duo, its first foldable	$1,999, ships Oct 23
Sept 9	Harvey raises late-stage round	$550M at $15.5B
Sept 9	Anthropic researcher Jacob Coxon resigns with a public warning	~76M views overnight
Sept 10	Anthropic publishes its fourth threat intelligence report	154 pages, 7 harm areas
Sept 10	DeepSeek releases V4.1-Flash	90.6 on Terminal-Bench 2.1

Science had a week too

Two research results would have led any normal news cycle. Google Research and HHMI Janelia published the complete connectome of a male fruit fly's central nervous system, charting more than 166,000 neurons and roughly 125 million synaptic connections, the largest brain ever mapped neuron by neuron. The dataset, published in Cell, lets researchers compare male and female wiring directly, and within days developers were piping the simulated fly brain into Doom and Super Mario 64.

Meanwhile Fei-Fei Li's World Labs debuted Atlas, a world model that builds persistent, explorable 3D environments from a single photo and natively handles text, images, video and 3D data. It generates up to 1440p video with pixel-level camera control, and outputs depth maps, point clouds and Gaussian splats. Atlas is early access only for now, but it is the clearest statement yet of Li's "spatial intelligence" thesis: models that reason about the physical world, not just describe it.

Apple folds, finally

At its September 9 "Surprise and Shine" event, Apple introduced the iPhone Duo, its first foldable: a passport-shaped device with a 5.4-inch outer display that opens to a 7.6-inch inner screen, starting at $1,999 and shipping October 23. Apple is years behind Samsung on folding hardware, and it is characteristically unbothered about that. The bet is that arriving with a mature hinge and an iPad-class canvas matters more than arriving first.

The warnings arrived in the same news cycle

The most-read AI post of the week was not a launch. Jacob Coxon, a 27-year-old researcher who spent three years on pretraining at OpenAI and then Anthropic, resigned with a public warning that the labs are "gambling with our lives" in a race dynamic, and that superintelligent systems carry "a risk of causing human extinction." His seven-part thread drew roughly 76 million views overnight, and colleagues inside Anthropic publicly echoed the concern.

A day later, Anthropic published its fourth threat intelligence report, a 154-page account of misuse it detected and disrupted between December 2025 and August 2026: state-linked cyber espionage, an automated fake-news operation, surveillance tooling aimed at dissidents, scams, weapons-adjacent research and model distillation. It is the most detailed public accounting any lab has given of what people actually try to do with frontier models, released in the same week the industry celebrated its biggest capability jump.

The application layer moved the same week

Products absorbed the frontier at record speed. ChatSlide launched ChatSlide 3, the biggest release of its AI presentation maker to date: slide generation on its Ultimate plan now runs on GPT-6 Astra, days after the model's API opened, and its new World Lab experiment turns an uploaded photo into an explorable 3D world you can drop into a deck. The gap between "frontier model ships" and "frontier model powers a consumer workflow" has collapsed from quarters to days.

That compression is the real story of the week. In ten days the industry declared AGI, priced it, funded it, productized it and warned against it, all at once. Whatever September 2026 turns out to have been the beginning of, it will be the week people point back to.

Cover image: AI-generated illustration.