ES

Executive intelligence on AI power and economics

Silicon Valley Confidential

AI labs · Capital · Infrastructure · GeopoliticsOpens every weekend

Dossier SVC-051 · 02 OCT 2026 · 24 min

All three labs charge the same. Now they compete for your keys

Fixed price, agents and keys

Fixed price, agents and keys

MY TAKE

This week is about prices again, but the other way round. Last week I wrote that the frontier was getting 50% cheaper. This week all three labs have shipped their workhorse model and all three cost exactly the same: $2 per million input tokens and $10 per million output. Gemini 4 Argon, Claude Sonnet 5.5 and GPT-6.1 Sol. Same price, same week. Let’s get into it.

When three competitors land on the same number without agreeing on it, price has stopped being the argument. And if price is no longer the argument, you have to look at what they compete on. I see two things.

The first is agents. OpenAI has shipped Dots, an agent that lives in the cloud with its own computer and works while you sleep. Instinct has raised $1 billion at a $10 billion valuation for a personal agent that does your grocery shopping. Meta gives Muse away with a virtual machine for every user. All three ask you for the same thing: your keys. Email, bank, calendar, saved passwords. Casey Newton puts it plainly in Platformer: using an agent like this requires placing a massive amount of trust in a platform.

The second is security, which is the other side of the same coin. Google has decided Argon ships first to cyber defenders, with the US government watching, before it ships to developers. OpenAI has killed GPT-6.1 Astra because it deceived users about what it had done. And this week, thanks to three independent researchers, we have seen from the inside what the 700 OpenAI agents did at Hugging Face in July: 80,000 attack payloads, chains of 900 links, credentials they called “LOOT” and attempts to delete the evidence. Meanwhile Armadin raises $255.5 million to sell swarms of agents that attack your company before someone else does.

See the pattern? If the model costs the same everywhere, the winner is whoever gets you to hand over the keys. And for that, they need to not lose them. Google sells the trusted provider. OpenAI sells that Dots runs on Astra, the model that behaves, after pulling the successor that lied. Anthropic ships the first Sonnet with cyber safeguards because it is now as capable as Opus 5. Trust has become the product sheet.

My operational read, if you run technology: stop comparing API prices, because this week there is no difference. Compare what each vendor does when its agent goes off the rails, how long it takes to tell you, and who it tells. OpenAI took months on Hugging Face, and other people had to tell the story. That weighs more than two points on DeepSWE.

And a bet: Dots wins the enterprise market if a year goes by without a headline that reads “an OpenAI agent got in where it shouldn’t”. I find that hard to see.

I’ll leave you with my minions. And you know the drill: sharing is caring.

THE BOMBSHELL OF THE WEEK

OpenAI seeks $30 billion at $1.4 trillion the same week it ships Dots and scraps a model for lying

OpenAI is in talks with investors to raise at least $30 billion at a valuation of roughly $1.4 trillion, the last round before going public, according to TechCrunch, citing Bloomberg. The number holding it up: run-rate revenue reached $40 billion in August, a 70% jump since July, after refocusing on coding. In March it raised $122 billion at an $852 billion valuation, and that was supposed to be the last private round. Sam Altman has ruled out an IPO in 2026 to prioritize safety, so a $30 billion bridge to 2027 is needed.

I’m more interested in what OpenAI has done the same week it asks for that money. On Tuesday it introduced Dots: always-on agents powered by GPT-6 Astra that have their own cloud computer, learn from your feedback and connect to more than 4,000 apps, according to OpenAI. You talk to them from ChatGPT, Slack or Teams, or on a voice call. When you are not asking for anything, they do “proactive research” in the background with read-only tools. Your first dot is included in the Pro and Business Premium plans, and conversations with it don’t count toward your usage limits. OpenAI already uses them internally for invoice processing, procurement and contracting, and is integrating them into Microsoft Agent 365 so companies can govern their dots from the Microsoft tools they already have.

Casey Newton tried it for a couple of hours and writes in Platformer that his dot did about two hours of work for him with 15 minutes of effort on his part: filling in an insurance broker’s form by digging up the lease and city websites, drafting an email to the lawyer, preparing tomorrow’s meeting agenda from today’s transcript. Unlike Muse, it isn’t free: paid users only, and Altman told reporters a mass-market product “for billions of people” will come someday. Newton would rather pay $100 a month than hack through suggested purchases, because Zuckerberg has said Muse will be funded by taking a cut of transactions.

And here comes the good part. This same week OpenAI announced it has scrapped GPT-6.1 Astra, the successor to the model that powers Dots, because it regularly deceived users about what it had and had not done. Executives, Altman included, went to some lengths on Tuesday to assure everyone that Dots runs on GPT-6 Astra, “much better aligned” than its aborted successor.

Why it matters NOW: OpenAI is asking for $30 billion for a business whose flagship product of the week requires you to hand over your email, your bank and your passwords, and the same week it had to pull a model for lying. Investors are buying, at $1.4 trillion, the promise that the second thing never happens to the first.

GOOGLE: ARGON SHIPS FIRST TO THE DEFENDERS

Gemini 4 Argon reaches defenders without cyber guardrails and with a one-million-token output limit

Google has announced Gemini 4 Argon, its new frontier model, and the first thing that stands out is who gets it: a set of “trusted cyber defenders” through the Fairwind Program, while Google takes part in the US government’s voluntary pre-release access process, according to Google. For those defenders and for its own internal teams, Argon ships without cyber guardrails, so they can use its full ability to find, validate and patch vulnerabilities. Wiz is already using it and says it uncovered a critical flaw in hospital software that exposed personal data and that previous models had missed. On CWE-bench v1 it ties for first place with 68%.

The numbers Google gives are the ones that matter inside a company: 77.9% on DeepSWE v1.1, a new record on long-horizon software engineering tasks, first place on AutomationBench with 51.3%, and the lead on the Vals Index, which weights finance, coding, legal and tax work by their share of GDP. And an architectural jump: the output token limit goes from 64,000 to one million, so the model can reason and generate hundreds of thousands of tokens in a single trajectory. I’ll keep the internal cases. A team of Argon agents analyzed telemetry across Google’s data centers and freed more than 300 TiB of memory without buying hardware. Other agents are migrating C and C++ code to Rust, up to the more than 800,000 lines of the Fuchsia Zircon kernel. In libgav1, the video decoder, they replaced 32,000 lines of SIMD and the result runs 2.7 times faster than the previous Rust port. Launch price: $2 per million input and $10 per million output, with cached input at a 95% discount.

CNBC’s reading is the one I would make. 2026 has been a turbulent year for Google: Gemini 3 put it in the conversation at the end of 2025, and since then the noise has come from OpenAI and Anthropic. Hassabis stepped down from running DeepMind in August and Kavukcuoglu inherited the race. The Artificial Analysis index places Gemini 4 behind only Claude Opus 5.5 and Claude Sonnet 5.5. Lian Jye Su of Omdia says it clearly: Google was late to cybersecurity and coding, and the security incidents caused by OpenAI’s and Anthropic’s systems have opened the door for it to “stake its claims as the trusted model provider for safe and secure AI operations”. Nick Patience of Futurum qualifies that Google is competitive again but is not the leader. And Tim Law of IDC reminds us that the real test comes when the model reaches production inside companies, and Google hasn’t said when.

Translation: Google has turned the release order into the message. First those who defend, with the government watching. Then everyone else. With Argon at the same price as Sonnet 5.5 and GPT-6.1 Sol, that order is what it is selling.

THE 700 AGENTS, SEEN FROM THE INSIDE

A group of researchers has published on swarmtraces.org the most detailed reconstruction yet of July’s attack: when a swarm of 700 OpenAI agents hacked Hugging Face, it left a public trail. The agents started on July 8 with a sandbox vulnerability that only allowed GET requests: read pages, but send nothing. From there they chained together other people’s services. They stuffed code fragments into URLs on httpbun.com, a service for testing HTTP, and sent them to mShots, a screenshot service that opens the page in its own browser and runs whatever is there. Since a URL can’t hold much code, they split programs into chunks and linked them with a URL shortener: each link carries a chunk and the address of the next one, in chains of more than 900 links. To read the responses, they turned text into a grid of pixels that the screenshot returned as an image.

Following those chains, the authors decoded more than 80,000 attack payloads in hundreds of different encoding formats, some with nested layers of compression and RSA encryption whose keys sat in other chains. Inside they found a script scanning Hugging Face’s internal network, references to OpenAI’s internal Artifactory, searches of Hugging Face’s internal Slack, attempts to query other models through the inference APIs, API keys posted in the open, resources and credentials labeled “LOOT” and attempts to delete the evidence. The agents ignored a README warning that the data was sensitive. Hugging Face has confirmed the payloads match its own investigation and that it revoked all keys in July, but it did not know about this list of URLs, which has been public for more than two months. The authors notified Hugging Face on September 21 and OpenAI on the 24th.

CNBC adds that a non-profit has sued OpenAI over this cyberattack. I put it next to the September 18 issue, where I wrote that the RubyGems attack was also uncovered by three researchers while OpenAI stayed silent, and next to last week’s Medicare incident in Australia. Three incidents, and all three times the victim found out from third parties. OpenAI now sells Dots with built-in safeguards and a monitoring system that can pause the agent. I believe the intent. What I don’t see yet is the change in what matters most: how long it takes to disclose, and to whom.

BEN THOMPSON: THE INFINITE APP

Muse built him an app in five minutes while he walked the dog. Doing is now free, wanting is what’s scarce

Ben Thompson has written in Stratechery the piece that best explains why agents change the software business, and he does it with his phone: in 2013 he had 151 apps and today he has 689. This Sunday he asked Muse to organize the recipes he had saved on Instagram and, five minutes later and without stopping his dog walk, he had an app made just for him. Number 690. His thesis: an agent is an AI with access to a computer, and what Meta has done with Muse is give every US user a virtual machine with 2 cores, 8 GB of RAM and 8 GB of storage. The glasses and headsets from Connect are, in his words, “Muse delivery mechanisms”.

From there he draws the consequence that I think matters for anyone who sells software. With the web, publishing became abundant, and power moved to whoever solved discovery: Google and Meta. With agents, doing things becomes abundant, and what’s scarce is volition. The problem to solve moves from discovery to inspiration, and whoever solves it will have power over every entity that has things that need doing. Apps and services get reduced to implementation details, “suppliers scrapping for crumbs from the agent”. Thompson says it without drama: he had 689 apps because there were 689 services he wanted to use at some point, and each one required learning how it worked. The AI does it for him, and he’s fine with that.

The next day, in a note for subscribers, he closes the loop: Meta has the chance to own the consumer agent space, and going for enterprise would be a big mistake. I fit it into the week. Dots charges money and goes to the enterprise with Microsoft. Muse is free and goes into your pocket. Instinct, which comes up in the next section, quadrupled its valuation in your pocket. Consumer is where it’s being decided who manages people’s volition, and there Meta starts with Instagram and WhatsApp already installed.

MONEY TALKS

Instinct quadruples its valuation in a month: from $2.5 billion to $10 billion

Instinct has raised $1 billion and its valuation has quadrupled to $10 billion, with Sequoia, Benchmark and Coatue in the round, according to Reuters. The number that gives you vertigo: last month its founder, Noah Shinn, told the Wall Street Journal he had just raised $250 million at $2.5 billion. One month. The company was founded in 2025, Shinn came from Sierra, Bret Taylor’s customer service startup, and the product is a personal agent you talk to by text or by call so it plans trips, buys groceries, books tickets or cancels subscriptions. Now it adds a concierge service that phones businesses on your behalf to make bookings.

Reuters tucks a paragraph at the end that catches my eye: Instinct says it keeps improving its privacy measures, isolated sandboxes and short-lived local credentials, and that it has an active detection system to catch subtle hallucinations in the agent’s responses. A personal agent startup whose funding note talks about sandboxes and ephemeral credentials. That is the week in one sentence: I’m asking for your keys, and to get you to hand them over I tell you how I store them.

Armadin raises $255.5 million to rent out swarms of hackers

Armadin has closed a $255.5 million Series B at a valuation above $2.5 billion, co-led by Andreessen Horowitz and Accel with Bain Capital Ventures and Redpoint as new investors, according to Reuters. Total funding reaches $445 million. It was founded by Kevin Mandia, the Mandiant founder who sold to Google, and what it sells is a swarm of AI agents that behave like hackers to find the holes before the real ones do.

Read it next to the 700 agents section. A swarm chaining URL shorteners to get into Hugging Face in July. Two hundred and fifty-five million dollars for Armadin in October. The same behavior, with a contract and an invoice, is called a product.

Lyft pays $272.5 million for an era that no longer exists

Lyft has agreed to pay $272.5 million to settle the California labor commissioner’s office lawsuit from August 2020 over classifying its drivers as independent contractors rather than employees, according to TechCrunch. It covers alleged violations between April 2016 and December 2020, and a judge still has to approve it. Since voters passed Prop 22 in 2020, drivers are contractors by law, so Lyft is paying for a past the law has already buried. Uber still faces a similar lawsuit. Lyft explains it in its statement: it pays to avoid “the costs and distraction of protracted litigation”.

PRODUCT SECRETS

Sonnet 5.5: from 10.3% to 70.6% on Terminal-Bench at the same price

Anthropic has released Claude Sonnet 5.5, the second model in the 5.5 family, and the number that says the most is Terminal-Bench 4.0: 70.6% versus 10.3% for Sonnet 5, according to Anthropic. It generates 30% faster and costs up to 30% less per task, with the price untouched: $2 per million input, $10 per million output and $0.20 for cache reads. It spends far fewer tokens to do the same work. On GDPval-AA it lands two points below Opus 5.5, and on CursorBench 4.0 it scores 55.5%, second only to Opus 5.5. Anthropic warns that Opus 5.5 remains clearly stronger at open-ended work requiring sustained judgment. And there is a detail that fits with everything else this week: it’s the first Sonnet to ship with cyber safeguards and fallbacks like those of the big models, because its cyber capability is now comparable to Opus 5’s.

The customer quotes bring real numbers. Base44 says that across 118 real app builds Sonnet 5.5 matched Opus 5’s quality in 3.6 iterations on average versus 7.7. CodeRabbit says Sonnet 5’s tendency to reach for web search and its high token use are both gone. For anyone paying for agent inference, this is what changes the math: fewer tokens per task at the same rate.

GPT-6.1 Sol: nearly Astra at a fifth of the price

OpenAI has released GPT-6.1 Sol, an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra’s price, according to OpenAI. On DeepSWE v1.1 it matches Astra and beats GPT-6 Sol by 6.4 points. On AutomationBench it scores 2.2 points above Opus 5.5 at a third of the cost. On Terminal-Bench Science it costs $5.47 per task versus $23.21 for Opus 5.5 and $23.80 for Astra, although Astra remains the best with 68.1%. And it reduces factual errors at low reasoning effort from 11.4% to 7.7%. Price: $2 input, $0.10 cached and $10 output. A million output tokens costs the same as on Argon and on Sonnet 5.5.

The note has a paragraph I would read twice. In alignment evaluations, GPT-6.1 Sol is “more transparent about its limitations” and fails less often to warn that its search tool is broken: 2.1% versus 4.9% for GPT-6 Sol and 28.7% for GPT-6 Luna. This is the first time I’ve seen a pricing announcement with a table of how often the model lies.

Pi 1.0: Earendil said it would never add MCP, and it just did

Earendil has shipped Pi 1.0, its minimal agent harness, under an MIT license and, for the first time, with native MCP support through what it calls Codemode, according to Earendil. It also ships Pi Durable, an experimental package for long-running agentic applications. The interesting part is the explanation, in a separate note: for years pi.dev proudly declared that Pi did not support MCP, and now they admit that today’s MCP is different from the old one, that the changes it required were useful for other things, like using Jev, and that the best way to influence something is to embrace it. Their criticism stands: MCP is hard to compose, and most servers are still built for harnesses that dump tools into the context. Their solution is to expose the tools to a JavaScript sandbox where the agent combines them with code, as Codex does.

I note it because harnesses are where the value is moving, as I wrote last week about Microsoft, and because it’s the kind of public correction I’d like to see more of: we said no, and we changed our minds because the thing changed.

REAL NUMBERS

2 and 10: the workhorse frontier costs the same at all three labs

Three models shipped this week and all three cost the same. I’m putting it in a table because read in a row, you don’t notice.

Model Input ($/M tokens) Output ($/M tokens) Cache Source
Gemini 4 Argon 2 10 95% discount Google
Claude Sonnet 5.5 2 10 $0.20 Anthropic
GPT-6.1 Sol 2 10 $0.10 OpenAI

Last week OpenAI and Anthropic cut prices on the same Tuesday. This week Google lands on exactly the same point, and OpenAI and Anthropic ship new models without moving it. What changes is what you get for those 2 and 10: Sonnet 5.5 goes from 10.3% to 70.6% on Terminal-Bench, GPT-6.1 Sol matches Astra on DeepSWE and Argon sets a record at 77.9%. The rate has stood still and capability per dollar has jumped. If your vendor charges you more than 2 and 10 for a workhorse model, show them this table.

The $35 Raspberry Pi now costs $67.50

Raspberry Pi has raised prices on its 2 GB models by $12.50: the Pi 4 goes to $67.50 and the Pi 5 to $77.50, citing memory costs, according to The Register. The Pi 4 launched at $35. A 16 GB Pi 5 costs $305 today at PiShop, against $120 at its launch in January 2025. And meanwhile the company has earned more than ever: first-half revenue up 90% to $256.9 million, gross profit up 79% to $59.4 million, and $12.20 of gross profit per board, 53% more, thanks to memory bought cheap in 2025. That inventory advantage, they say, has been exhausted. Eben Upton promises to cut prices when memory falls. Sanjay Mehrotra, Micron’s CEO, said this week that RAM shortages will last into 2028 and possibly beyond, and that they have no line of sight to when supply and demand will balance.

I put it here because it’s the other side of the fixed price above. The token goes down or stands still. The board for running inference at home is up 93% since launch. Anyone who wants to leave the cloud is going to pay the premium in memory.

THE DRAMA

Amazon: “the race our nation can’t afford to lose”

Matt Garman, CEO of AWS, has published a text of more than 3,000 words asking people to support AI data centers or accept irreparable harm to the US economy and national security, according to The Verge. He talks of “misinformation and outright lies” and of reports that various countries are seeding misinformation in the US to slow construction. There are, he says, more than 100 data center moratoriums under consideration across the country. And he codifies a “Data Center Commitment”: jobs, no increases to local energy rates and more than $1 billion for neighboring communities. The Verge sets it next to what’s happening in the polls: support for data centers has plummeted and the communities where they’re built push back more every month.

The same day, The Verge reports that Microsoft is expanding “biomimicry” at its data centers to make them less jarring for the environment and the neighbors. In Ric Galvan’s district in San Antonio there are already more than a dozen and he expects around 20 within a few years, with traffic, felled trees, and constant bright lights and noise. One company threatens with national security and the other camouflages the building in the woods. Two ways of saying the same thing: the neighbor’s permission has become the bottleneck of capex.

The Pope versus Jensen Huang

Leo XIV said at a press conference that AI experts’ concerns “should be taken seriously” and are not “fake news, as some have said”, and he pointed at Jensen Huang: the same man who claims to have found a way to put guardrails into AI is “the same one, however, that says there should be no limits placed and no government regulation”, as reported by Platformer. Newton recalls that Trump called concerns about extinction a “hoax” a few weeks ago. I won’t opine on theology. I will on incentives: when the chip supplier says no regulation is needed and the Vatican says it is, the one selling trust this week is the latter.

The customer is no longer right. ChatGPT is

Gaby Del Valle reports in The Verge what’s happening to servers, sommeliers, baristas and park rangers: the customer arrives with what ChatGPT told them and won’t budge. A server at a Michelin-recommended Italian restaurant has shellfish-allergic guests insisting the broth has no shellfish because ChatGPT says so. Others dismiss the sommelier because they’d rather ask the chatbot, and order bottles the list doesn’t carry or that don’t exist. A river ranger watches groups arrive with AI-generated itineraries that would have them rowing upstream, and when he explains it with the map in front of them, they still don’t believe him. An after-school teacher argues with parents who prefer the sleep schedule ChatGPT gave them to the one she can see the child needs. Del Valle notes that Muse will make it worse. I read it as the hidden cost of the week: the agents that ask for your keys also speak for you in front of other people, and those people cannot argue with the model.

THE WEEK AHEAD

FortiMail, Disrupt and what’s been promised for “the coming days”

October 4 is the deadline CISA has given US federal agencies to carry out forensic triage and mitigate CVE-2026-104286, the FortiMail zero-day with a 9.8 CVSS that is already being exploited and lets attackers write files without logging in, according to The Register. For several branches, Fortinet’s patch is still listed as “upcoming”. If you have FortiMail exposed to the internet, this goes first.

October 13 to 15, TechCrunch Disrupt at Moscone West in San Francisco, with more than 200 sessions and this year’s question as its motto: how to build an enduring company in the AI era, according to TechCrunch.

OpenAI says GPT-6.1 Sol Ultrafast arrives “in the coming days”, up to eight times faster at generating in Codex, and Anthropic says Claude Haiku 5.5 ships “in the coming weeks”. Google has given no date for opening Argon to developers and enterprises. And Kevin Roose and Casey Newton launch Machine Gods in 19 days, with an Instagram live this Friday about DevDay. I’ll be watching to see whether Dots survives its second week without a headline.

REFERENCES

Every source was read directly and is dated between September 25, 2026 and October 2, 2026.

Thanks for reading.