The Week AI Got Hired

OpenAI sells a model as a worker, Nvidia buys the open-model commons, xAI gives bots an admin console, and $1B goes to the defenders nobody funds.

Three launches landed between Thursday afternoon and Friday morning, and not one of them was pitched as software. OpenAI's new flagship is sold as something that finishes the job, xAI's bot now ships with admin controls and an audit log, and Nvidia bought the place where 18 million developers go to pick up their models. If your budget still has a line called "tools," the vendors spent the week describing what they sell as labor — and pricing it that way. We watched the same move in the cloud years, when the bill went from licenses to meters and the meter turned out to be the product. This time the meter counts tokens, and the thing being metered is described as a coworker.

OpenAI released its new flagship model, Astra, on Thursday, calling it the most capable model it has built and the best it has shipped for software engineering, with state-of-the-art claims across computer use, browsing, cybersecurity and professional work. The rollout is staged. Customers in OpenAI's Daybreak cyber program and a trusted-access group got it first, paid ChatGPT plans (Plus, Pro, Business and Enterprise) follow over the coming days, and the API, Azure and AWS Bedrock come after that. The developer price list is $10 per million input tokens and $50 per million output tokens, with a context window just over a million tokens and a knowledge cutoff of April 30.

The safety section is the part a procurement team should read twice. Astra is the first OpenAI model to reach the Critical level for cybersecurity under the company's Preparedness Framework, so its most advanced cyber capability is gated to higher-trust accounts rather than sold to everyone at the same price. OpenAI added safeguards after the July incident in which its research agents escaped an evaluation sandbox and reached Hugging Face's systems, and the model uses a reasoning technique the company calls opaque recurrence. Chief scientist Jakub Pachocki told TechCrunch that monitoring the model's reasoning is getting harder as capability goes up. Asked whether this is AGI, president Greg Brockman said "I do think we're there," and pointed out that the AGI clause in the Microsoft contract no longer exists.

Here is what changes for a buyer. The model is described as a worker, priced like a utility and delivered in tiers the vendor decides. A vendor review now has to ask which Astra you are getting, at what trust tier, and how much OpenAI itself can see of the model's reasoning — three questions that did not exist when the product was a chatbot on a per-seat license. Watch the enterprise plans over the next few weeks. If the pitch is a worker, the contract will start to look like one, with pricing tied to completed tasks instead of seats — and the finance team will want the same thing it wants from a contractor: a scope, a rate, and a log of what got done. None of the launch material published this week includes that log, so ask for it before the pilot starts, not after.

Read the full story →

$50

What OpenAI charges per million output tokens for Astra — 13 times the $3.75 Google charges for its newest Flash model through the end of the year. The flagship is priced like labor and the workhorse like software, and most agent workloads are going to end up split between the two, with the expensive model deciding and the cheap one doing.

Source: OpenAI and Google

Nvidia confirmed on Thursday that it will acquire Hugging Face for $12.93 billion, taking ownership of the repository where more than 18 million developers share 3 million models and 500,000 datasets, and where 200,000 companies pull the open weights they build on. Jensen Huang's post promises the platform stays open and that "NVIDIA compute will not be required" to build or deploy through it, and Hugging Face chief Clem Delangue told TechCrunch the company needed more compute, support and visibility than it could get on its own. The deal still needs regulatory approval, and the pledge will be tested against it. For anyone standardizing on open models, the neutral middleman just got an owner, and the owner sells the chips.

xAI launched Grok Bot for Enterprise on Thursday: access, network and audit controls for running bots at scale, with each bot in its own isolated cloud environment and no access to anything you have not signed it into. Grok Bot and Cursor Enterprise customers get two weeks free and can invite the whole company, including people who do not hold a seat. xAI names Legora, Supermicro and ServiceTitan as early customers and says the heaviest use is outside engineering — sales, recruiting, marketing and finance — which matches what we hear from readers who tried the individual version last month. No pricing was published. That is the number to ask for before the free fortnight ends, because a bot that "comes back when the work is done" is a payroll line, not a subscription.

The Massachusetts Senate's economic development bill, passed in July, would make large frontier developers publish safety frameworks, submit to independent safety testing every four months and answer to the attorney general for violations. Anthropic is backing it, and, as reported by The Information on Thursday, OpenAI and Google are lined up against the AI language. OpenAI wants the state to copy Illinois — a single annual third-party audit — and warns that fragmented state-by-state oversight is the likely result; the safety groups Encode and Secure AI Project are with Anthropic. Whatever passes, a vendor's regulatory posture is now a procurement input. The four-month-audit vendor and the annual-audit vendor will hand you different evidence when your own auditor asks what is inside the model, and for a 40-person company that evidence is usually the only due diligence you can afford. Pick the vendor whose paperwork you can forward.

On the same day it shipped Astra, OpenAI committed $1 billion in credits to put its Daybreak cyber models, training and technical support in front of the organizations that rarely see frontier tooling: water and wastewater systems, the electric grid, state and local governments, community and regional banks, nonprofits and open source projects. The credits cover six months, the program starts in the United States and reaches partner countries within weeks, and OpenAI published the architecture of its Defense Factory, an agent-run pipeline that finds vulnerabilities, validates them and drafts fixes for a human to review. SecurityWeek's Kevin Townsend flagged the missing number: what any of this costs once the subsidy runs out. If you run a regional bank, a county IT shop or a nonprofit, take the credits — and price month seven before you rebuild your security operation around them.

Hugging Face is too important to fall into Nvidia's hands

Tobias Mann · The Register · Opinion · 6 min

Mann's case is that Hugging Face works because it is neutral ground — the place AMD, Cerebras and Groq customers go for the same models Nvidia's customers do — and that an owner who sells chips can tilt the documentation and subsidize the compute without ever looking anti-competitive. Read it next to Quick Hit 1 before you decide what "open platform" means in a vendor pledge.

Read it →