top story → Google DeepMind releases Gemini 4 Argon at $2 and $10 per million tokens, Meta ships Muse Wonder and Anthropic gives METR employee-like access

THE SUNDAY

What the people building AI said this week

Sunday 18 October 2026 · · covers Sun 11 Oct to Sat 17 Oct · ≈ 10 min

Lead story · Models

Google DeepMind releases Gemini 4 Argon at $2 and $10 per million tokens

Google DeepMind made Gemini 4 Argon available on Tuesday through the API in AI Studio, to AI Ultra subscribers in Antigravity, and in the Gemini app. Launch pricing is $2 per million input tokens and $10 per million output, with cached input at 95% off, and Willison notes the price is set to double on 1 December. Google says Argon ranks first on the Vals Index, and that teams using it in Antigravity raised agentic code submissions by more than 35% while cutting rollbacks by nearly a third. Zvi says it is too early to know whether the model lives up to its benchmarks.

Google DeepMind · Simon Willison · Zvi · Tue 13 to Thu 15

##Models4 stories

Meta releases Muse Wonder, and its Full tier scores 56 on the Intelligence Index

Meta shipped Muse Wonder and Muse Wonder Full on Monday. Latent Space says Wonder is on the Meta Model API and in Muse Code, Full is for Muse Code subscribers only, and Wonder is priced at $4 and $20 per million input and output tokens with a 1M context window. Artificial Analysis scores Full at 56 and Wonder at 52 at max effort, up from 48 for Muse Spark 1.3, with Claude Opus 5.5 still ahead at 58. Zvi says Meta's benchmark presentation looked suspect and that it is too early to judge.

Latent Space · Artificial Analysis · Simon Willison · Zvi · Mon 12 to Thu 15

Intelligence Index, selected models

Effort setting as listed by Artificial Analysis. Bars scaled to 65. Muse Wonder Full, Muse Wonder and Muse Spark Open are new rows.

ModelWeightsIndex
Claude Opus 5.5 (max)closed58
Claude Sonnet 5.5 (max)closed56
Muse Wonder Full (max)closed56
Claude Fable 5.1 (max)closed53
GPT-6 Astra (max)closed53
Gemini 4 Argon (high)closed53
Muse Wonder (max)closed52
GPT-6.1 Sol (max)closed52
Muse Spark Openopen39

closed weights  open weights  · Artificial Analysis leaderboard, checked Sun 18 Oct · Muse Spark Open's 39 is as Latent Space reports it

Reflection puts Beam on Hugging Face under Apache 2.0

Beam has 501B parameters with 23B active, a 1M token context, and checkpoints in BF16, FP8 and NVFP4. It ships with its own Rust based harness and day-0 support in transformers, vLLM and SGLang.

Hugging Face · Fri 16

Also The same Import AI covers Humanity's Sixth Sense, a benchmark of 522 questions on what a visual scene implies · Hugging Face hosts NVIDIA's 2.8 trillion token Indic pretraining set and open data generation pipeline · IBM's Granite Switch 4.1 serves chat, decisions and speech from one endpoint.

##Agents & tooling4 stories

Anthropic builds Scion to replace Git after merged changes grew from 2,500 to over 30,000 a day

Willison reports Anthropic's engineering post: merged code changes averaged around 2,500 a day in May 2026 and passed 30,000 a day in September. The main branch history alone is almost 100 GB, and CI runs about 25 million checkouts a day. A team of seven finished the migration from Git in five months, and Anthropic credits Google's Piper and Meta's Sapling as prior art. Zvi's AI #190 calls Scion Anthropic's internal replacement for GitHub.

Anthropic · Simon Willison · Zvi · Tue 13 to Thu 15

Drew Breunig names a new context failure mode, Context Inertia

It is when a model commits to a hypothesis or subtask and keeps pursuing it after the evidence says to stop. He points to Anthropic's Memento harness, where each agent writes a short note after a fixed timeout and a fresh agent takes over with only the note.

Drew Breunig · Wed 14

Geoffrey Huntley fixes agent drift with a pre-commit hook

When an agent skips his diagnostics library, he fails the commit and injects a prompt telling it to use the library. He says that took the agent to ten out of ten, which updating AGENTS.md could not guarantee.

Simon Willison · Tue 13

Also Python 3.15 is out, and Willison singles out explicit lazy imports from PEP 810 and drops support for 3.10 · Addy Osmani and colleagues, quoted by Willison, say review on the Claude Code team now runs continuously and Claude has reviewed most changes before a human looks · Lean Rada's document.currentScript trick replaces an inline element with a value when the page loads.

##Security4 stories

Anthropic gives METR employee-like access and names SecureBio and Accenture as evaluators

Anthropic has onboarded six METR employees with company hardware, office space and internal tools roughly on par with its own risk assessment teams. METR will publish its analysis of Anthropic's risk levels, incidents and practices, and Anthropic cannot veto it but can require redactions in predefined categories. Anthropic is also working with SecureBio and Accenture, and is in early talks with others including Hugging Face.

Anthropic · Thu 15

Zvi covers OpenAI firing three safety researchers

Per Zvi, Tomek Korbak, Mikita Balesni and Jasmine Wang were let go for what OpenAI called mishandling sensitive company information. The three published an open letter saying they believe the reasons were pretextual, and OpenAI says there was a further breach of trust that it did not detail.

Zvi · Mon 12

Zvi calls a leaked Anthropic draft AI bill unacceptable

He describes it as similar to the FRONTIER Act with stronger preemption of state laws, stronger transparency and weaker enforcement. He says he has not yet read the full bill and that Anthropic is purportedly shopping it around.

Zvi · Sat 17

Goodfire argues neural networks are already clean

Latent Space leads with Nick Cammarata's essay claiming that many known cases of polysemanticity and superposition give way to further agentic analysis. It calls the piece a rare story on interpretability and alignment.

Latent Space · Fri 16

Also Zvi's AI #190 reports a group used ChatGPT to build a full-chain Chrome exploit and claimed a $250,000 bounty · Anthropic's midterms post samples about 250,000 US Claude.ai conversations: 5.75% touch politics, 0.51% are about elections, and about 0.01% ask whom to vote for.

##Money & compute2 stories

Anthropic tells investors its third quarter revenue was $16.8bn, up from $11.5bn

Willison summarises the FT report: revenue was about $4.8bn in the first quarter, $11.5bn in the second and $16.8bn in the third. The third quarter brought $1.70 of revenue for every $1 spent on computing power, with an adjusted operating margin of 11% before stock-based pay. The FT also says distillation by other labs made up 5% to 7% of revenue and was cut back materially in the third quarter. Willison notes Anthropic no longer seems to lean on annualized run-rate revenue with investors.

Simon Willison · Thu 15

Also Latent Space's Muse Wonder digest lists Artificial Analysis cost per index task: $4.52 for Wonder, $7.63 for Fable 5.1, $5.98 for Opus 5.5, $3.26 for Astra and $1.99 for Argon, and says Meta stock rose 3% on launch day.

##Opinion3 columns

Nathan Lambert · Sat 17

US open weights are back

Lambert calls Meta's Muse Spark Open the strongest American-made open model by quite some margin. With Reflection's Beam, he says the gap between the best American and Chinese models has roughly halved, though 20+ Chinese open-weight models would otherwise beat the best Western ones.

Meta shakes up the frontiers, open and closed

Zvi · Wed 14

The safety culture at the labs is broken

Zvi says Anthropic cultivates a safety culture, but that much of the company relies on muddling through. He argues OpenAI's problems run deeper and points to a growing view that the frontier needs pacing.

The Safety Culture at the Labs is Broken

Popcar via Simon Willison · Wed 14

AI decompilation is awesome and scary

Popcar describes a wave of AI decompilation that enables game mashups and a boom in preservation, often running in a browser through WebAssembly. He warns that anyone can decompile software, so a paid app can be open-sourced against its owner's will.

AI is mass decompiling games. That's both awesome and scary.

##Audio

Dyna-2: general purpose physical agents with hour-long autonomy

Dyna's Jason Ma talks about a world-action model pre-trained on one million hours of human video and a robot that folded restaurant napkins at a 99.4% success rate over 24 hours. Dwarkesh had nothing new, so this is the one episode.

Latent Space · Tue 13

##Back pagesworth re-reading

##And finally

Willison tells Claude to make a game fun, and it works

He asked Opus 5.5 Max for a three.js game about appreciating the differences between two generated trees. After 3 hours and 7 minutes it returned a spot-the-difference game with music and achievements, and he admits make sure it is fun works uncomfortably well.

Tree Appreciator, Wed 14

##Wire

##Colophon

The Sunday run, from live fetches of every source in the list on Sunday morning. Feeds where they exist (DeepMind's feed needed decompressing), the page itself for Anthropic and Artificial Analysis. Anything already linked in issue 7 was dropped. Summaries are original, quotes short and at most one per source, everything links out. The window is Sun 11 Oct to Sat 17 Oct; nothing new turned up for Sun 11.

Three caveats. Nineteen items were filtered or skipped: eleven OpenAI customer, partner or education posts (Itaú, Stack Overflow, Benefit Street Partners, Sullivan & Cromwell, SentinelOne, Amp, Third Moment Research, Block, Higgsfield, the Rosalind Academic Plan and the students post), Anthropic's Singapore office opening, Zvi's monthly roundup, which is mostly off topic, and a Latent Space lunch-show pilot. Five Willison release notes and demos were left out to stay under the cap of 12 per source. OpenAI's items are listed from feed summaries and the Hugging Face wire items, apart from Muse Spark Open and Beam, from their titles only.

Sources polled

18 checked · 7 quiet · 43 wire items. Counts include items relayed through a source.

SourceKindKept
Simon Willison's Weblogblog12
Latent Spacenewsletter8
Hugging Facelab blog7
Don't Worry About the Vase (Zvi)newsletter5
DeepMindlab5
Anthropic Newslab, page2
Interconnectsnewsletter1
OpenAI Newslab1
Import AInewsletter1
Drew Breunigblog1
Artificial Analysistracker, pageupdated
Armin Ronacher, Hamel Husain, Eugene Yan, SemiAnalysis, Ahead of AI, One Useful Thing, Dwarkeshvariousquiet

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

the sunday rebase · issue 8 · sun 18 oct 2026 · 0 7 * * 0 · next run sun 25 oct 2026, 07:00 cest