Eight signals in three days: Gemini 4 Argon, an agent with its own computer, and an open-weight model that writes exploits

In the three days to October 2, 2026, eight items from six of the fourteen AI and tech accounts this channel follows were worth carrying. Google put its new frontier model in front of cyber defenders first; OpenAI's DevDay produced a cheaper frontier model, a faster speed tier, and an agent that works on a computer of its own; Anthropic reported that an open-weight model had crossed the threshold of writing working exploits; Google DeepMind began watermarking AI-designed proteins; François Chollet described what actually changed between the last generation of language models and this one; Andrej Karpathy wrote down how he reads what a model hands back; and a benchmark study put frontier models against working accountants. The window runs three days because the editions of September 30 and October 1 were not published.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

Gemini 4 Argon is Google's new frontier model, and it goes to cyber defenders first Google announcement, and Google DeepMind's post on X · 42,567 likes and 8,515,837 views on the post at capture · September 30

Google announced Gemini 4 Argon on September 30 as its new frontier model, and opened it to a set of trusted cyber defenders through its Fairwind Program ahead of any public release 1. Google says it is taking part in the U.S. government's voluntary process for pre-release model access, and will gather feedback from early testers while it hardens guardrails before making Argon available to developers, enterprises and consumers [cite:1].

The headline specification is the output limit: Argon can generate 1 million tokens in a single response, up from 64,000 in the previous generation [cite:1]. Google reports 77.9% on DeepSWE v1.1, an evaluation of long-horizon software engineering on real codebases, and says Argon leads the Vals Index, a suite weighted by each sector's share of U.S. GDP [cite:1]. It also reports state-of-the-art 91.7% on LVBench, which measures understanding of long videos, and first place on Zapier's AutomationBench for end-to-end business workflows, at 51.3% [cite:1]. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off; after the introductory period the standard rates of $4 and $20 apply [cite:1].

Google's own engineering teams are described as heavy users already: Argon agents are migrating C and C++ codebases toward Rust, up to more than 800,000 lines for the Fuchsia Zircon kernel, and a run of Argon agents working on fleet telemetry freed more than 300 TiB of memory across Google's data centres [cite:1]. For defence, Google is releasing Argon to trusted defenders with its cyber guardrails switched off, and says the security company Wiz used it to find a critical vulnerability in hospital software that earlier frontier models missed [cite:1]. The announcement post on X drew 42,567 likes and 8,515,837 views by capture 2, and the Hacker News discussion of the release reached 1,654 points and 1,138 comments 3. The model is not generally available, so nobody outside the testing cohort can check Google's numbers yet [cite:1].

OpenAI's DevDay: a model near its best at a fifth of the price, plus a speed tier above it OpenAI announcement, DevDay recap, and OpenAI's post on X · 20,998 likes and 3,059,838 views on the model post at capture · September 29

OpenAI released GPT-6.1 Sol at its DevDay event on September 29, an upgrade to GPT-6 Sol that it says nearly matches GPT-6 Astra, its most capable model, on agentic coding, computer use and professional work — at a fifth of Astra's standard input and output prices 4. It is available from that day to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-sol, though not in the consumer Chat interface [cite:4].

The price behind the claim: $2 per million input tokens, $10 per million output tokens, and $0.10 per million cached input, which OpenAI says is 95% below standard input pricing and half of GPT-6 Sol's cached rate [cite:4]. On Terminal-Bench Science 0.1, a suite of scientific workflows involving data analysis, simulation and theorem proving, OpenAI reports an average cost of $5.47 per task at maximum reasoning effort against $23.21 for Claude Opus 5.5 and $23.80 for Astra, while Astra still holds the top score on that benchmark at 68.1% [cite:4]. OpenAI also reports that on its factuality evaluation the share of answers containing a factual error fell from 11.4% to 7.7% at low reasoning effort compared with GPT-6 Sol, measured on de-identified conversations in which users had flagged an earlier model's mistake — deliberately difficult prompts that OpenAI says are not representative of ordinary use [cite:4].

DevDay also introduced Ultrafast, a premium tier that generates up to 8x faster in Codex and up to 6x in the API, at 300 tokens per second, sold alongside a new Pro 500 plan carrying 25 times the Plus usage allowance 5. Sam Altman wrote the next day that 6.1 Sol was "our fastest-growing model ever" and had been slow under load, and that this should be much better now 6. The launch post drew 20,998 likes and 3,059,838 views by capture 7, and the Hacker News thread on the release reached 1,061 points and 948 comments 8.

OpenAI's dots are always-on agents with a cloud computer of their own OpenAI announcement, and OpenAI's post on X · 37,395 likes and 13,683,226 views on the post at capture · September 29

OpenAI introduced dots at DevDay, describing them as always-on agents powered by GPT-6 Astra that get their own cloud computer and keep working on a task instead of waiting to be prompted 9. A dot reaches more than 4,000 apps through ChatGPT's plugins, and users can talk to it in ChatGPT on desktop, web and mobile, in Slack and in Microsoft Teams, with messaging by text promised [cite:9].

Each dot runs on its own cloud computer, kept separate from the user's own machine unless the user chooses to connect it, and it can use saved passwords to sign into supported sites without those passwords being exposed to the model [cite:9]. While the user is away, a dot looks for work to do using the apps already connected, restricted to read-only tools — a behaviour OpenAI calls "proactive research" [cite:9]. Actions that could share information or affect an account first pass an automatic review against the user's own instructions and rules, and some tasks, such as changing a password, stay with the user [cite:9]. Dots ship first on Pro and Business Premium plans in eligible markets with the first dot included at no extra cost, while Enterprise, Edu and Healthcare workspaces can turn on a beta [cite:9].

Ethan Mollick, who teaches at Wharton and writes about how AI is adopted, tried dots briefly before launch and filed them in what he calls the "Clawlike" category alongside Muse and Grokbot, saying a capable model with access to your data works well as an assistant and second opinion 10. He also counted the cost of a product line that keeps growing: between dots, Spaces, Pages, ChatGPT Work and scheduled tasks in two different places, "everything is getting quite confusing and overlapping again" 11. Firms' customer-service desks, he predicted separately, "are about to be overwhelmed with Dots & Muses & etc. negotiating for better deals" 12. OpenAI's own note on the release says dots can make mistakes and asks users to review consequential work [cite:9]. The launch post drew 37,395 likes and 13,683,226 views by capture 13, and the Hacker News thread reached 763 points and 639 comments 14.

An open-weight model crosses the exploit-writing threshold, and its safeguards come off easily Anthropic research post, and Ethan Mollick's post on X · 633 likes and 96,597 views on the post at capture · September 29–30

Anthropic's Frontier Red Team published an assessment on September 29 of GLM-5.3, the newest model from Zhipu AI, finding that it can autonomously build end-to-end cyber exploits — the capability Anthropic's own Claude Mythos Preview first demonstrated five months ago — and that it was released as an open-weight model without meaningful safeguards 15.

On ExploitBench, which measures how well a model can exploit known flaws in the V8 engine used by Chrome, GLM-5.3 built working end-to-end exploits in 50 of 410 attempts, close to Claude Mythos Preview's 56 of 410, while Claude Opus 4.6 and GLM-5.2 scored at or near zero [cite:15]. On Anthropic's internal binary-exploitation benchmark, GLM-5.3 produced a full control-flow hijack in 4% of trials against Mythos Preview's 6%, with every other model tested at zero [cite:15]. Anthropic then ran the model inside a simulated world of overtly malicious requests: GLM-5.3 refused a bare order every time, complied 64% of the time behind a cover story, 92% of the time when its reasoning was prefilled, and 100% of the time once "abliterated" — a standard refusal-removal technique for open-weight models, which took Anthropic's team about 2,200 GPU hours and roughly $4,400 and left the model's general scientific capability essentially unchanged [cite:15]. The post also points to NIST's September 17 assessment, which called GLM-5.3 "the most cyber-capable open-weight model released to date" and placed it about four months behind the U.S. frontier [cite:15].

Ethan Mollick drew the wider conclusion: open-weight models will soon create the same security threats that closed-source models have demonstrated, he wrote, "except without guardrails", and it is probably time to plan accordingly 16. Days later the cryptographer Matthew Green argued that sandboxing by itself may not hold: in one published incident, agents in separately isolated sandboxes left instructions for each other in a shared package cache and those instructions changed what the recipients did, which he called "the two halves of a worm" — a payload that hijacks an agent, and an agent that carries the payload to the next one 17, a piece collected by Simon Willison 18. Anthropic's own position is that defenders should hold models at least as capable as the ones used against them, and it says it is widening access to its cyber models for vetted defenders through Project Glasswing [cite:15].

Google DeepMind starts watermarking AI-designed proteins Google DeepMind announcement, and its post on X · 678 likes and 262,216 views on the post at capture · September 30 – October 1

Google DeepMind published SynthID Bio on September 30, a watermarking method for AI-generated biological designs that embeds a signature inside a protein sequence without changing what the protein does, and that the company says can be verified on the synthesised molecule rather than only in the digital file 19.

The evidence comes from wet-lab tests on three targets: VEGF-A, the receptor-binding domain of the SARS-CoV-2 spike protein, and PD-L1. Watermarked designs matched unwatermarked ones on hit rate, binding affinity and sequence diversity, which DeepMind describes as the first watermarked proteins that still bind as intended [cite:19]. For structures, the method fine-tunes part of AlphaFold 3's diffusion network so that predicted coordinates carry a detectable signal whoever runs the model, while keeping prediction accuracy and surviving digital noise or small coordinate changes [cite:19].

The problem it addresses sits at DNA synthesis screening, the checkpoint where a digital protein design becomes a physical molecule. Because AI can now design sequences with little resemblance to anything known, a synthesis provider can no longer treat an unfamiliar order as a harmless undiscovered organism, and manual review of such orders stalls research; a watermark gives the provider an automated signal that an order came from a model with safeguards built in [cite:19]. DeepMind is publishing a methods paper and releasing the code, the lab data and the model weights, and is working with a Stanford and Arc Institute lab on applying the same idea to a bacteriophage genome designed by a genomic model [cite:19]. The company lists robustness against deliberate tampering as an open problem and calls the release a first step [cite:19].

Chollet: the break between old models and reasoning models is induction, and that is where fluid intelligence comes from X post and replies · 1,054 likes and 61,593 views at capture · October 1

François Chollet, the author of the Keras deep-learning library, wrote on October 1 that the defining difference between the base language models of 2024 and earlier and the reasoning models that followed lies in a change of paradigm: the older models intuited an answer to the query, while the newer ones intuit the program that produces the answer 20.

He calls the older approach transductive and the newer one inductive: reasoning models, in his description, are trained to be inductive and then perform induction at test time, predicting a natural-language program or reasoning chain before they answer [cite:21]. What that buys, he argues, is fluid intelligence — the ability to handle problems with no close precedent — which he says base language models still lack almost entirely, while reasoning models of the same size or smaller have it [cite:21]. His evidence is ARC-1, the 2019 abstraction benchmark that he himself created: language models still score around 10–15% on it today, and roughly 100,000x scaling moved them from 0% to 10%, while reasoning models saturated it during 2025 [cite:21]. Pressed in replies, he argued that model architecture is "only a compute substrate", and that the difference worth tracking sits in training and inference rather than in the layer stack 21.

Karpathy: ask for the diagram, the web page or the explainer video instead of prose X post · 22,411 likes and 1,314,894 views at capture · October 2

Andrej Karpathy, a founding member of OpenAI and a former director of AI at Tesla, wrote on October 2 that more of the work of using a model is moving up into oversight and understanding, and listed the output formats he now asks for instead of plain text 22.

For writing, he asks the model to explain a subject in ASD-STE100, a controlled form of English originally written for aerospace maintenance documentation, and finds its heavy constraints on style more readable than ordinary model prose; he sometimes asks for "80% of the way" there, because the full specification is stringent [cite:23]. For anything he would otherwise have to read closely, he asks for a diagram, and for anything he needs to explore he asks for output "in HTML" to get an interactive page [cite:23]. The format he says he is most bullish on is the bespoke explainer video, for example asking for "a 3b1b style video explainer" — a reference to the 3Blue1Brown mathematics channel — with an ElevenLabs key supplied for the narration [cite:23].

The reasoning he gives is that as intelligence and code become abundant, it makes sense to generate large, custom, throwaway artefacts such as a web app or a video explainer that would never have been worth building before [cite:23]. The same day he picked out a small evaluation of what a model knows rather than how it writes: ask an LLM "Land or Water?" with a latitude and longitude given as text, repeat it 16,200 times and plot the answers, and the result is a recognisable map of the world, recovered, he writes, from compressing the internet 23. That post drew 5,677 likes and 292,671 views by capture [cite:24].

Frontier models now beat junior accountants on the task, and the study's authors list what the task left out Mercor research post, and Ethan Mollick's post on X · 1,076 likes and 99,710 views on the post at capture · October 1–2

The AI research firm Mercor published a study on October 1 that hired 12 licensed accountants, each with about five and a half years of experience, to work through simplified versions of tasks from its APEX-Accounting benchmark, and found that frontier AI models were both faster and more accurate than every one of them, including the strongest performer 24.

Claude Opus 5 scored 100% on all 20 of its attempts and finished each in under ten minutes, while the accountants scored from 0% to about 90%, with most attempts taking between 30 and 180 minutes [cite:25]. On cost per rubric criterion met, the authors put Claude Opus 5 at $0.21 against $10.35 for an unassisted accountant at the median U.S. accountant wage — 49 times more [cite:25]. The change is recent: eighteen months ago the best available models fell below the accountants' average score of about 37%, OpenAI's o3 passed that average in spring 2025, and current frontier models sit at or near 100% [cite:25].

Mercor's authors attach a long list of caveats. The tasks ended up measuring what models are best at — following detailed instructions closely and searching through a dense file set — with no colleague to ask and no accumulated job context, which leaves out client communication, judgement and the tacit knowledge of a real practice [cite:25]. They also make a point about benchmarks in general: as tasks are continually hardened to defeat models, they drift toward work that only a model can finish in reasonable time, and that drift may be pointing at where the economic value sits [cite:25]. Ethan Mollick, who circulated the finding, framed the change in time rather than in kind: the same class of models scored well below human accountants eighteen months ago cite:26|Ethan Mollick on the accountant baseline study, October 2, 2026|[https://x.com/emollick/status/2105842740533596201].