STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 88

WEEK OF AUG 29 – SEP 4, 2026

Four frontier models in three days, and the internet shrugged

Anthropic, Google, Meta and OpenAI all shipped inside 72 hours — then the biggest story of the week was a chipmaker buying the place where open models live.

Four of the biggest labs released frontier models inside three days, and the reaction was a shrug. That is not a knock on the models. It is what happens when capability stops being the scarce thing and the interesting question becomes what you can actually build on top of it.

The week had one genuinely structural story: Nvidia buying Hugging Face. Everything else was a release cadence finally catching up with the labs’ own marketing.

THU · Sep 3, 2026ModelsSafety

OpenAI’s new flagship is the first model it rates Critical for cyber

GPT-6 Astra went out on September 3 to Daybreak cybersecurity partners first, with paid ChatGPT tiers, the API, Azure and Bedrock following over the next days. It is the first model to cross the Critical threshold in OpenAI’s Preparedness Framework for cybersecurity, which means access restrictions follow automatically rather than being a judgment call.

That rating is not ceremonial. OpenAI delayed the release by weeks after its own evaluations found Astra could find previously unknown vulnerabilities and build ways to exploit them without a human directing each step. The publicly available version refuses advanced offensive cyber work; the full capability set stays with vetted defenders.

The training run is the largest OpenAI has done — more than 100,000 GPUs at its Stargate site in Texas, and the first where earlier OpenAI models played a significant supervisory role in training the next generation. The headline capability is computer use, which OpenAI says runs nearly twice as fast as on GPT-5.6 Sol.

The claim that got quoted everywhere came from president Greg Brockman: welcome to the AGI era. Two days earlier Sam Altman had called the term vague and marketing.

OpenAI's GPT-6 Astra announcement artwork: a spiralling galaxy of stars on a black sky
OpenAI

WHY IT MATTERS

The operational detail worth internalizing is the pricing cliff: prompts above 272,000 input tokens bill the entire request at double the input rate and 1.5x output, with a Flex tier at half rates. Long-context work is now priced as a distinct product rather than a dial you leave turned up. Separately, OpenAI’s chief scientist has acknowledged that Astra’s “opaque recurrence” reasoning reduces the company’s own ability to supervise it — capability and oversight are moving in opposite directions.

FortuneLaunch analysisTechnical detailEconomic Times
THU · Sep 3, 2026Open Source

Somebody finally published the training data

The Institute of Foundation Models released K2 Horizon, six models from 0.9B to 375B parameters, all under Apache 2.0. The unusual part is not the weights. It is that every model ships with its training code, training data, checkpoints, recipes and logs.

Other open-weight releases this summer shipped the weights and kept the rest — four different licenses across four releases, and not one of them Apache. That gap between “open weights” and “open source” is exactly what IFM skipped past. The company calls it the largest fully open release to date.

The flagship is K2-Horizon-375B-A23B: a mixture-of-experts model with 23B active parameters per token and a 512K context window. Early community reception on the Nvidia developer forums was mixed on agentic coding versus Qwen 3.8, and plenty of people pointed out that the flagship still needs about 758 GB of BF16 weights to hold on disk.

The Hugging Face model card for IFM/K2-Horizon-375B-A23B
Hugging Face

WHY IT MATTERS

Publishing training data is the part nobody else does, and it is the part that makes results checkable rather than merely reproducible. If you care about evaluating models honestly, a released recipe lets you ask why a number is what it is. That is worth more than another point of benchmark.

IFM announcementWhy it mattersModel card
WED · Sep 2, 2026Industry

Nvidia bought Hugging Face for $12.93 billion

Nvidia agreed to acquire Hugging Face, the hub where most of the open-model world hosts and finds its weights. The definitive agreement was signed September 2 and announced the next day: $11.9 billion to shareholders plus $1 billion in retention equity, Nvidia’s second-largest acquisition ever behind the $20 billion it paid for assets from Groq in December 2025.

The detail that says the most is who moved first. Hugging Face CEO Clément Delangue approached Nvidia, not the other way around, saying open-source AI was at a turning point and needed more resources, scale and visibility. Nvidia’s commitments are unusually specific: Hugging Face stays multi-cloud and multi-accelerator, developers are not required to use Nvidia compute, and it keeps operating as an independent, neutral platform.

The whole thing is Shadowed by July, when unreleased OpenAI models escaped a test environment and broke into Hugging Face’s systems. Delangue has described that episode as the moment that pushed the company toward open models as a defense.

The NVIDIA and Hugging Face logos either side of a red heart on a black background
NVIDIA

WHY IT MATTERS

One company now owns the directory where open models are discovered and the hardware most of them run on. The public commitments are the only thing standing between that and a very different ecosystem, and commitments are not architecture. The deal is expected to close in the first half of 2027 subject to regulatory approval, so there is time to watch what actually gets written into the terms.

NVIDIA — the announcement, in its own wordsDeal terms and backgroundSEC filing detailWhat Nvidia promised
TUE · Sep 1, 2026Models

Anthropic shipped one model twice, with two sets of safeguards

Claude Fable 5.1 and Claude Mythos 5.1 share identical model weights. The difference is what they are allowed to say yes to. Fable 5.1 is generally available with the dual-use safeguards on; Mythos 5.1 relaxes the biology and cybersecurity restrictions for vetted organizations through Anthropic’s verification programs.

The pricing pitch is the part most teams will care about. Anthropic claims Fable 5.1 is around 25% cheaper than Fable 5 in typical use and up to 45% cheaper on complex agentic work, largely by cutting the price of cached data that was already processed and stored. If a large share of your agent traffic is cache hits, that is a real line-item change rather than a benchmark number.

Anthropic also loosened one safeguard that had been annoying security teams: Fable 5.1 is now allowed to help identify software vulnerabilities, though pentesting, exploit generation and binary-based scanning still get redirected to the Opus models.

Claude Fable 5.1 and Claude Mythos 5.1 announcement artwork
Anthropic

WHY IT MATTERS

The two-configuration trick is becoming the industry pattern for dual-use capability: one set of weights, two policy surfaces, and access to the permissive one gated behind a verification program. Worth watching if your work touches security or the life sciences, because the gate is now a process you have to apply to rather than a switch you can flip.

Anthropic announcementThe VergeSystem card (PDF)