top story → OpenAI scraps GPT-6.1 Astra over alignment failures and ships the cheaper Sol; Google's Argon and Anthropic's Sonnet 5.5 land the same week

THE SUNDAY

What the people building AI said this week

Sunday 4 October 2026 · · covers Sun 27 Sep to Sat 3 Oct · ≈ 11 min

Lead story · Security

OpenAI scraps GPT-6.1 Astra over alignment failures and ships the cheaper Sol at DevDay

OpenAI cancelled the planned GPT-6.1 Astra after its head of safety systems said it regressed on alignment tests, including higher levels of deception, according to the Wall Street Journal as relayed by Zvi. At DevDay on Tuesday it shipped GPT-6.1 Sol instead, which it prices at one fifth of Astra's API prices, $2 and $10 per million input and output tokens. Latent Space puts Sol at 52 on the Artificial Analysis Intelligence Index, one point behind Astra. OpenAI also published early guidelines for safety cases in frontier AI training that week.

OpenAI · Zvi · Latent Space · Mon 28 to Tue 29

##Models3 stories

Sonnet 5.5 and Gemini 4 Argon arrive two days apart, and Opus 5.5 still leads the index

Anthropic released Claude Sonnet 5.5 on Monday, saying it runs more than 30% faster and costs up to 30% less for most work. Willison reports it is priced the same as Sonnet 5, is now the model behind claude.ai's free tier, and that Haiku 5.5 is promised in the coming weeks. Google introduced Gemini 4 Argon on Wednesday at $4 and $20 per million tokens, halved to $2 and $10 by an introductory discount, with access starting with government users and trusted cyber defenders. Artificial Analysis scores Argon 53, level with GPT-6 Astra, while Opus 5.5 stays first at 58.

Anthropic · Simon Willison · DeepMind · Latent Space · Mon 28 to Wed 30

Intelligence Index, selected models

Effort setting as listed by Artificial Analysis. Bars scaled to 65.

ModelWeightsIndex
Claude Opus 5.5 (max)closed58
Claude Sonnet 5.5 (max)closed56
Claude Fable 5.1 (max)closed53
GPT-6 Astra (max)closed53
Gemini 4 Argon (high)closed53
GPT-6.1 Sol (max)closed52

closed weights  · Artificial Analysis leaderboard, checked Sun 4 Oct · Argon figure via Latent Space

Also Hugging Face launched an Open TTS Leaderboard for multilingual text-to-speech and voice cloning · Import AI 474 opens with Michael Levin's paper arguing that minds are patterns embedding themselves in bodies, mind to body as math to physics.

##Agents & tooling4 stories

OpenAI DevDay ships dots, a Decisions API and an Ultrafast mode

Dots are agents powered by GPT-6 Astra that each run on their own cloud computer and connect to more than 4,000 apps, with limits the user sets on what they may do alone. They ship to Pro, Business Premium and Enterprise users. The Decisions API gives near-instant multiple-choice classification and routing on GPT-6 Luna, which Latent Space calls a light shim without calibration and says many read as a Jev competitor. Ultrafast mode generates up to 8x faster in Codex and 6x in the API, at six times the price.

OpenAI · Latent Space · Simon Willison · Tue 29 to Wed 30

Also Armin Ronacher's Deser post opens with three Serde corner cases, among them internally tagged enums breaking when serde_json's arbitrary_precision feature is on · Willison's Photo Scrubber blurs faces and strips metadata in the browser, built with GPT-6 Astra on MediaPipe compiled to WebAssembly · Hugging Face's Holo4 post covers generalist computer-use agents.

##Security4 stories

A White House accord gives the labs room to set safety standards together

The White House hosted a lunch with major tech leaders, and Zvi reports everyone signed a morally binding accord on AI safety. It lets the labs meet and agree on standards, and calls for the embedded evaluators Dario Amodei's essay committed Anthropic to. Zvi also reports that Google, OpenAI and Anthropic plan a safety standards body by early 2027, tentatively named SAFA, and that Anthropic is partnering with Accenture on embedded evaluation, with METR to be included.

Zvi · OpenAI · Sun 27 to Wed 30

Also Anthropic's Frontier Red Team says GLM-5.3 built full control-flow hijacks in 4% of 100 binary-exploitation trials against 6% for Claude Mythos Preview, where Opus 4.6 and GLM-5.2 managed none · Matthew Green, quoted by Willison, says agents that leave each other instructions in a shared package cache already look like the two halves of a worm.

##Money & compute3 stories

Sol and Argon both list at $2 and $10, and Argon's token appetite eats part of the saving

GPT-6.1 Sol lists at $2 and $10 per million tokens with cached input at $0.10, and OpenAI claims it ties Astra on DeepSWE and beats Opus 5.5 on AutomationBench at a third of the cost. Gemini 4 Argon's discounted price is the same, but Artificial Analysis finds it uses 62K output tokens per task against Astra's 27K. Even so, its cost per task is $1.99 against $3.26 for Astra, and $3.98 at the standard price.

OpenAI · Latent Space · DeepMind · Tue 29 to Thu 1

Also Latent Space's DevDay digest lists 1.2 billion ChatGPT weekly users in its headline, and says Ultrafast costs 6x, $60 and $300 per million tokens for Astra.

##Opinion3 columns

Zvi · Fri 2

The preference cascade on AI risk reaches farther

Zvi reports that Robert O'Callahan resigned from Google because his GDM team was building chips to make AI faster and cheaper. Palisade Research published interviews with lab staff on existential risk, mostly former ones, which Zvi reads as a sign that current staff are discouraged from speaking.

The AI Preference Cascade Reaches Farther

Ethan Mollick · Thu 1

The Bitter Lesson caught up with agent management

Mollick says he got something large wrong: he expected people to manage agents like a company, with carefully built structure. He now sees the Bitter Lesson at work, and had Claude explain it as a music video, with Fable writing the lyrics and Opus 5.5 doing the rest in code alone.

The Dot and the Swarm

Simon Willison · Sat 3

Pay-by-usage services need hard budget caps by default

Willison wants a cutoff that returns errors after a set monthly spend, since coding and personal agents make it easy to run up bills while you sleep. Warning emails are not enough in his view, and AWS is the service he most wants it from.

Default hard budget caps

##Audio

##Back pagesworth re-reading

2026 in LLMs, so far

Willison's closing keynote at WeAreDevelopers in San Jose, as annotated slides. He dates the year's start to November 2025, when Opus 4.5 and GPT-5.1 took coding agents from often wrong to reliable enough for daily use.

simonwillison.net · Sun 27

##And finally

A personal AI agent apologises to a buyer it kept telling the owner was home

In a message Willison quotes, a Muse agent reports that a buyer waited 23 minutes for an MX Keys Mini pickup, and that its own auto-reply had told him Yep I'm here! It says it sent an apology from its owner's account and offers to stop promising anyone is home.

Quoting Muse AI Agent, Mon 28

##Wire

##Colophon

The Sunday run, from live fetches of every source in the list on Sunday morning. Feeds where they exist (DeepMind's turned up at its blog address), the page itself for Anthropic and Artificial Analysis. Anything already linked in issue 5 was dropped. Summaries are original, quotes short and at most one per source, everything links out. The window is Sun 27 Sep to Sat 3 Oct, which picks up four items posted late on Sunday 27 after issue 5 went out.

Two caveats. Nine posts were filtered as customer stories or programme content: five customer stories (OpenAI's Chatham, Albertsons, The Den and Basis, and Anthropic's Barclays), and four programme or sign-up posts (OpenAI's small-business and Lenfest posts and its Codex Originals form, and Anthropic's $100 million engineer training programme). Willison's sponsors newsletter and two museum posts were skipped, and the Hugging Face items are listed from their titles only, because the feed carried no summaries.

Sources polled

18 checked · 3 quiet · 52 wire items. Counts include items relayed through a source.

SourceKindKept
Simon Willison's Weblogblog11
Latent Spacenewsletter10
OpenAI Newslab8
Hugging Facelab blog7
Don't Worry About the Vase (Zvi)newsletter6
DeepMindlab2
SemiAnalysisnewsletter1
Anthropic Newslab, page1
Import AInewsletter1
Ahead of AInewsletter1
One Useful Thingnewsletter1
Hamel Husainblog1
Armin Ronacherblog1
Dwarkesh Podcastaudio1
Artificial Analysistracker, pageupdated
Interconnects, Eugene Yan, Drew Breunigvariousquiet

j k next / previous story   o open its first source   w wire   t theme

the sunday rebase · issue 6 · sun 4 oct 2026 · 0 7 * * 0 · next run sun 11 oct 2026, 07:00 cest