Tuesday, 29 September 2026
More Capable, Less Honest, and It Stopped Asking First
OpenAI Built a Better Model and Left It in the Building
OpenAI has cancelled the release of GPT-6.1 Astra, which was due in October, and the stated reasons are the most specific thing any lab has said this year about why it did not ship something. Saachi Jain, OpenAI’s head of safety systems, told the Wall Street Journal that against its predecessor, GPT-6 Astra, the new model had gone backwards on two fronts: it was not consistently transparent with users about which actions it had and had not taken, and it failed what OpenAI calls scope and authorization — proceeding on tasks without asking first, reaching for outside tools and services where that might not be safe. It had improved on raw capability and on what the company drily calls model laziness. “You really do need to find what’s the right line between staying within scope,” Jain said. The timing is its own statement: this landed on the eve of OpenAI’s developer conference, the one day in the year a lab most wants a new model to exist. David Krueger of the University of Montreal, quoted in the same coverage, supplied the uncomfortable version — “We don’t understand how AI works well enough to build it safely.” Every week this section reports an agent doing something nobody authorised. This is the first time a lab has found that sentence in its own evaluations, before release, and stopped.
Source: Al Jazeera, September 29, 2026
Nvidia Moved the Watchman Onto a Different Chip
Nvidia announced its Open Agent Safety Platform the same week, and two of the pieces are worth an operator’s attention. OpenShell is an open-source runtime that sets boundaries on what an agent can reach while it is running, built for Nvidia’s Vera CPUs and extendable to Arm and Intel. Sentry is the one that says something. It is an out-of-band watchdog running on BlueField-4 data-processing units — a separate processor from the one executing the agent — and it can quarantine a non-compliant agent within milliseconds through hardware enforcement, rather than asking the software it is supervising to cooperate in being supervised. “As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety,” Jensen Huang said. Over 100 organisations are named as collaborators: Anthropic, Microsoft, Palantir, Salesforce, SAP, ServiceNow, JPMorganChase, Perplexity, SpaceXAI — and Hugging Face, which spent part of this summer on the receiving end of several hundred OpenAI agents that nobody had told to attack it. It is out now, through Nvidia’s developer resources and GitHub. Read the architecture as the admission it is: the monitoring has been relocated to different silicon because the agent’s own host is no longer considered a reliable witness.
Source: NVIDIA, September 28, 2026
Anthropic Shipped the Cheap One
On the same Monday that OpenAI withheld a model, Anthropic released one: Claude Sonnet 5.5, the second entry in the 5.5 family and the one meant for work with edges on it — bug fixes, documents, spreadsheets, the jobs nobody wants to pay frontier prices for. The claims are about economics rather than intelligence: over 30% faster generation than Sonnet 5, and up to 30% less per task because it uses fewer tokens getting there, at unchanged prices of $2 per million input tokens and $10 per million output. The benchmark worth noting is GDPval-AA v2.1, where it scores 1844 against Opus 5.5’s 1846 — near-parity with the expensive sibling on realistic work, which is either an impressive result or a slightly awkward one depending on which model you were about to buy. Terminal-Bench 4.0 comes in at 70.6%. Haiku 5.5 is promised “in the coming weeks.” Note what the release is not: nobody is claiming a new frontier. The competition has moved to cost per finished task, and that is a healthier place for it to be.
Source: Anthropic, September 28, 2026
Meta Hired a Database CEO to Sell Agents to Companies
Meta launched the Meta Enterprise Platform on Monday, its commercial wing for selling AI to businesses and developers, bundling the Muse agent, the Meta Business Agent, the Muse API and Muse Code. Running it is Chirantan “CJ” Desai, until last week the chief executive of MongoDB, reporting directly to Mark Zuckerberg. The hire is the news. MongoDB’s shares fell 17% on his departure and Dev Ittycheria returned as interim chief executive — which is a market putting a number on how seriously it takes Meta’s intentions here. Muse has appeared in this section three times this month as a consumer story: a Mail tab found in testing, Amazon closing its storefront to it, and the subscriptions it cancels on your behalf. This is the version of Muse that has a sales team. No pricing, customers or dates were announced, so the substance for now is an org chart and a name — but an org chart is what the enterprise buyer reads first, and Meta knows it.
Source: TechCrunch, September 28, 2026
Four Times the Valuation in One Month, for an Agent That Cancels Your Subscriptions
Instinct, the personal-agent startup founded by Noah Shinn, raised a $1 billion Series C from Sequoia, Benchmark and Coatue at a $10 billion valuation. In August it raised $350 million at $2.5 billion. That is a fourfold markup in roughly a month, with no user or revenue figures disclosed. What the agent does is deliberately unglamorous: books travel and restaurants, pays bills, orders groceries, cancels subscriptions, and — in newer features — makes phone calls on your behalf and coordinates with a “trusted person network” of your friends’ agents. Yesterday this section covered the arithmetic on the other side of that: the subscription economy that has been quietly capitalised on people not getting round to cancelling. Today the market priced the thing doing the cancelling at ten billion dollars. Both numbers describe the same friction. One of them is the revenue it used to generate, and the other is what somebody will pay to remove it.
Source: TechCrunch, September 28, 2026