The frontier gets 50% cheaper. The model is no longer the moat

MY TAKE

This week is about prices. And about what happens when the price stops being the news. Let’s get into it.

First fact. On Tuesday OpenAI shipped GPT-6 Sol and Luna with the API 50% cheaper than their predecessors. The same Tuesday, Anthropic shipped Claude Opus 5.5, which costs 40% less than Opus 5 and performs like Fable 5.1 on almost everything. Two frontier price cuts on the same day. That is what I call a market.

Second fact. Ben Thompson has written the best analysis of the year on why this is happening: model capability is now good enough for most customers, and when something is good enough the moat moves somewhere else. To the harness, to the user touchpoint, to the data. He proves it with a case I love. Anthropic tried to make customers accept data retention to use Fable, customers said no, and in Fable 5.1 the clause disappeared. The customer chose on something other than performance. That is where the model moat ended.

Third fact. Meta has put Muse, a personal agent that is not even a month old, at the center of everything. 500,000 users in the first week according to The Information, “millions” according to Zuckerberg, and up to $145 billion of capex this year that needs a story. And watch out, because Thompson says Muse is the best personal agent he has tried without its model being the best. Same thing again: good enough.

What do the three facts have in common?

That the fight has moved. From the smartest model to whoever keeps the user and their context. Microsoft has understood it and puts everything into one Copilot app that bills by usage. Meta is trying with an agent that lives inside your accounts and a pair of 100-gram glasses.

If you run technology, my operational read: renegotiate your API prices now, because they just halved and your vendor will not call to tell you. Do not marry your product to a model. And look at where your data lives, because that is what will cost you to move, far more than the model.

If I had to bet: in six months Sonnet 5.5 and Luna are the default choice for almost every agent in production, and within a year this ends with everything at 50 cents per million tokens. The $20 Opus charges for output today belongs to another era.

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

THE BOMBSHELL OF THE WEEK

OpenAI and Anthropic cut the price of the frontier on the same Tuesday

OpenAI has introduced GPT-6 Sol and GPT-6 Luna, Astra’s cheaper siblings, and cut its API prices by 50% against GPT-5.6’s promotional pricing, according to its announcement. The numbers it gives are the kind that hurt a competitor: on AutomationBench, Sol at xhigh effort beats Claude Opus 5 at max effort at 9% of the cost per task. On DeepSWE v1.1, Sol lands 1.1 points behind Claude Fable 5 (68.8% versus 69.9%) at 80% lower cost. And Luna, the small one, matches Opus 5 and Fable 5 at medium effort while costing 93% and 96% less. Cache reads carry a 90% discount, and GitHub says the caching improvements have cut the tokens needing fresh processing by more than 50%.

Anthropic answered the same day with Claude Opus 5.5: it performs like Fable 5.1 on most work and costs 40% less than Opus 5 on typical workloads, according to Anthropic. Input tokens at $4 and output at $20 per million, 20% less than Opus 5, and cache reads, which make up most of the cost of agentic work, at $0.20, 60% less. It generates 30% faster. It is the first model Anthropic has published since it called for pacing the frontier, and external evaluators such as Frontier Design tested it before release.

What interests me in the two posts is what they confess without meaning to. OpenAI says its median researcher spends $600 a day on tokens and the 90th percentile spends $7,000. Anthropic says that “at these levels of capability, benchmark margins have become a less reliable guide to real-world differences” and that the gap between Opus 5.5 and Fable 5.1 is “narrower than these scores suggest”. Both labs are saying the same thing with a different accent: the model no longer sells just for being better.

Ben Thompson explains it in Stratechery with Christensen’s theory. While the product is not good enough, the integrated player wins. Once it is, the modular player wins and competition moves to speed, convenience and price. His proof is Fable’s data retention clause: Anthropic imposed it, customers would not play along, and in Fable 5.1 it was gone. “Pure capability no longer translates directly into a moat”. Microsoft, he says, saw it earlier: Nadella runs a multi-model harness that rotates GPT, Claude and open weights, and you can now pick a model in Copilot Cowork.

Translation: if the model is interchangeable, the infrastructure math changes. CNBC reports that Anthropic and OpenAI are now hunting for small data center deals, 20 to 30 megawatts, in the UK, the Nordics and the United States, on top of the gigawatt monsters. The reason is inference: training needs thousands of chips working together, serving can be spread across small clusters. JLL estimates inference will go from 9% of data center capacity in 2025 to 37% in 2030, overtaking training in 2027. We said it last week with Crusoe and its data centers on trucks: new infrastructure is being designed to serve models far more than to train them.

Why it matters NOW: if you build on these APIs, your bill just halved without you doing anything, and your vendor will not tell you. Renegotiate. And if you were picking models by benchmark, stop. Even the people who make the benchmarks are telling you they no longer discriminate.

META: MUSE AND A PAIR OF 100-GRAM GLASSES

Zuckerberg needs a story for $145 billion and he has found one

At Connect, Zuckerberg said “the centerpiece” of what Meta is building is Muse, a personal agent released less than a month ago, and that in a few years it will be “the personal superintelligence that billions of people are going to use”, as Casey Newton reports in Platformer. Muse is a consumer version of OpenClaw: you connect your Google Workspace, your banks and your services, and ask it to fill in forms, grab reservations or pay bills. For free. It is getting a voice, it is being built into the glasses’ operating system, and the Muse Charm, a sort of strapless watch for talking to your agent, ships later this year.

The numbers Newton puts next to it are the ones that matter. Zuckerberg said on stage that “millions” have tried Muse. The Information had reported the day before 500,000 users in the first week. Threads hit 100 million in five days. And Meta plans to spend up to $145 billion this year on infrastructure and capital expenditure, more than the $85 billion Reality Labs has lost since 2021. A single year’s capex exceeds all the metaverse losses. With that bill you need a story, and Muse is the story.

I am more persuaded by what Thompson writes in Stratechery: Muse is “by a significant margin, the best and most approachable personal agent product” he has tried, and its model, Muse Spark 1.3, is not state of the art. That is the warning for the labs. A personal agent you have poured your life into is far stickier than a chatbot, and you do not need the best model to build one. Your code is on GitHub and you switch agents in an afternoon. With your whole life inside Muse, switching costs a good deal more.

Amazon has already blocked Muse, and Thompson devotes another piece to explaining that Amazon’s physical-world investments are a moat against AI, and a third to Walmart as Meta’s way to wait out Amazon, to Expedia defending its middleware position, and to the question of where Google is. Aggregator against aggregator: if the agent shops for you, the store loses its shop window.

And the glasses. Newton, who has spent a decade trying headsets that end up in the garage, comes out impressed. The Meta VR Glasses weigh 100 grams, one sixth of the Vision Pro, at roughly a third of the price, with the computing in a separate puck that keeps them from overheating and three hours of battery. The display Meta calls 5K has under half the Vision Pro’s pixels and he did not notice in the demo. The usual questions are still there (for gaming, for work, for movies?), but as engineering it is the best he has tried. As he puts it, you cannot lose $85 billion on something without learning a thing or two.

OPENAI: ANOTHER INCIDENT DISCLOSED THREE MONTHS LATE

An agent broke into Australia’s Medicare portal looking for the price of a skin treatment

Australia’s prime minister, Anthony Albanese, has revealed that an OpenAI agent gained unauthorized access to a public-facing Medicare portal, read public and private files and wrote files to an internal server, as Platformer reports. The breach happened on June 18. The government found out by email on September 10. Albanese called Sam Altman to express “Australia’s extreme concern” and his disappointment that “it took the company way too long to inform the government”.

What makes this case special is explained by Transluce, the research lab that analysed the incident. The agent was not on a cybersecurity task. It was in a mundane internet research evaluation and broke into the portal to find the average cost of skin treatments in the Australian healthcare system. In Transluce’s words, malicious activity “can arise instrumentally to solve mundane tasks like information retrieval”.

Last week I covered the RubyGems attack that OpenAI never disclosed and three researchers uncovered. This week, Medicare, disclosed three months late and by the affected government. The incident-reporting framework OpenAI inaugurated ten days ago still does not cover what matters most: whom you tell and when. Reuters puts it next to a $6.4 billion security round, because the money is already pricing these incidents in. That is what the next section is about.

WASHINGTON JOINS X’S FIGHT WITH BRUSSELS

The Justice Department wants the first DSA fine annulled

The US government has applied to intervene in the challenges by Elon Musk, X Internet and X Holdings against the €120 million fine the European Commission imposed on X in December 2025, the first under the Digital Services Act, according to The Register. The Commission sanctioned the paid blue checkmarks, the advertising repository and the obstacles placed in front of researchers. Assistant Attorney General Brett Shumate accuses the Commission of having “inappropriately attempted to expand its regulatory authority to reach American companies not present or operating within its jurisdiction”, and warns: “We will not tolerate the European Commission engaging in regulatory overreach to try and control American engines of innovation”.

The technical detail is the one that will matter at the General Court. The Commission found Musk and his companies jointly and severally liable and calculated the fine on the worldwide turnover of the group he controls. Justice says that extends scrutiny to Musk himself as a private individual and to other companies of his, and that if it stands it will have “significant implications” for the large platforms with parents in the United States. Ted Cruz called the fine “an abomination” and Brendan Carr of the FCC said Europe is “fining a successful US tech company for being a successful US tech company”.

What you will not read anywhere else: the Commission has already accepted X’s plan to fix the advertising repository and researcher access. What remains standing is the €120 million and the way it was calculated. Washington is entering the case for the legal theory, which is what affects Meta, Google, Apple and Amazon. X is the excuse.

MONEY TALKS

Island is worth $6.4 billion because agents keep leaving their lane

Island, which makes enterprise browsers and security tools, has raised $400 million at a $6.4 billion valuation, more than 30% above its $4.8 billion of 2025, according to Reuters. Evolution Equity Partners leads, with Sequoia, Coatue, Insight and J.P. Morgan Growth returning. It has 1,000 employees and has doubled its recurring revenue every fiscal year since launching in 2022.

Reuters links the round to the Medicare incident of the same Thursday and to the agents that bypassed OpenAI’s controls in July. I read it the same way. Every agent that escapes is a sales argument for whoever sells the fence, and a security company doubling revenue every year while the labs argue about slowing down is how the market states its opinion.

Lightspeed halves its India fund and dedicates it to AI only

Lightspeed is raising $250 million for Lightspeed India Partners V, half the $500 million of its 2022 fund, and already has 80% committed, according to an investor letter seen by TechCrunch. It starts investing within two months, with an investment period of about two and a half years, and for the first time puts its India funds on the firm’s global cycle, as Accel did in August with $550 million for India inside a $3.5 billion raise.

The shift is in the thesis. The previous fund invested in whatever came along (quick commerce, audio, rooftop solar), this one only in AI, on the idea that AI will create more value in India than the internet did. Lightspeed manages more than $65 billion and raised $9 billion last December, so $250 million is loose change for the firm. What it tells me is something else. India has not produced a frontier lab and has attracted far less AI capital than the United States and China, and the bet is still the application layer, built on its pool of developers. A small, fast fund for a market where someone else supplies the model. It fits the week.

PRODUCT SECRETS

Microsoft puts chat, code and agents in one app and bills it by usage

Microsoft has unveiled the new Copilot app with three tabs, Home, Code and Autopilot, according to The Verge. Home merges Copilot Chat and Cowork with a daily dashboard of emails, meetings and Teams threads. Code lets any employee create an app, a dashboard or an automation and share it as an internal app hosted in the tenant, on the same technology as GitHub Copilot. Autopilot is Scout under a new name: a cloud agent with its own identity, memory, computer and workspace, which lives inside the tenant and which you @mention in Teams like a colleague.

The news for whoever pays is the billing model. Cowork, Code and Autopilot are billed by usage, including the models such as Astra and Fable running underneath, with an automatic model picker and a “FinOps for AI” so the admin does not get a shock at month end. Nadella calls it “a new OS for work” and Spataro compares it to what Office was for the PC.

I put it together with Thompson: Microsoft is the company that separated the harness from the model, and now it sells the harness. Models getting cheaper suits it perfectly, because it charges for the orchestration and the model is a cost that keeps falling. The Verge recalls that Copilot has failed in the consumer space against Claude and ChatGPT, and that the new app does not compete with Muse. It is “the AI built for work”. Home and Code reach the Frontier program in the coming weeks and Autopilot goes into private preview this month.

Gravity Linux lands accelerated Linux on the M4 Mac mini with agents

A fork of Asahi Linux called Gravity Linux has released an alpha that boots an accelerated desktop on the M4 Mac mini, according to The Register. The odd part is how. Cody Ho, formerly of Apple and OpenAI, and Niklas Sheth built the GPU driver in about a month with coding-agent loops that ran largely unattended. Asahi explicitly forbids generative AI in material contributions, and that is where the fork comes from. Gravity allows it with a clean-room policy: if your agent has looked at disassembly of an Apple component, you are “tainted” for that component and someone else has to write the implementation.

Asahi has lost Hector Martin, Asahi Lina and Alyssa Rosenzweig in a year and a half, and is only now starting on the M3. The Register is fair about it: this is a demonstration of what AI-assisted reverse engineering can do, and not a controlled experiment proving Asahi’s policy held it back. The demonstration is good enough for me.

Laya: the open answer to Jev comes out of India and runs seven times faster

Last week I covered the launch of Jev, TypeSafe’s model that generates no text and returns decisions with calibrated probabilities. This week Nandakishor Mukkunnoth, founder of ConvAI Innovations, has published Laya with a message in the title: he built the same thing a year ago, with an arXiv paper in March 2025 and weights on Hugging Face, and a well-funded lab has presented it as a brand-new breakthrough. Laya is his answer: open weights under Apache 2.0, 32.8 milliseconds per decision on one GPU (7.2 milliseconds per question batched), 6 to 8 times faster than Jev, and more than 100 languages.

On his own benchmarks it beats Jev on typed-decisions (0.766 versus 0.727), on AG News (0.950 versus 0.910) and above all on calibration, with an error of 0.081 versus 0.246. And this is what makes me take it seriously: he publishes where it loses. With more than 20 options it collapses (0.425 versus Jev’s 0.870 on Banking77) and without fine-tuning it is close to random. A model that tells you where it does not work is a model I can test.

Tesla delivers the Semi nine years later

Tesla has begun deliveries of the Semi this week, unveiled in 2017 and in volume production since April, according to The Verge. The factory next to Gigafactory Nevada can build 50,000 trucks a year. Customers: PepsiCo, DHL, US Foods and Einride, with an order for 500. What Tesla did not say: how many orders it has or what it costs. In 2017 it promised $150,000, and Electrek has reported it quotes the 500-mile version at $290,000. It consumes 1,700 watt-hours per mile fully loaded, and with diesel above $6.50 a gallon the operating math works. Musk was not at the event. He was at the state dinner with Trump and Xi.

REAL NUMBERS

Waymo has 4,000 robotaxis and 80% of them are in two states

Kirsten Korosec has done in TechCrunch the exercise I wanted to see: looking at where Waymo’s cars actually are. It has gone from 3 cities in September 2024 to 15, and does 500,000 paid rides a week. But of its roughly 4,000 robotaxis, 80% are in California and Texas. And Texas is where everything is happening: the fleet has grown 49% in three weeks, to 1,102 vehicles registered as of September 24.

The growth engine is a Chinese minivan. The Ojai, a modified Zeekr RT from Geely, is already a third of the Texas fleet. It arrives without Chinese connected-car electronics, gets outfitted at Waymo’s Arizona factory, and pays tariffs that eat the savings it was supposed to bring. Waymo absorbs them: MoffettNathanson estimates it will bring in 5,100 Ojai this year. The demo is Silicon Valley, the scale is Texas and the car is made by Geely.

FrontierSWE: Fable 5 dominates, but the benchmark is already archived

On FrontierSWE v1, 17 ultra-long-horizon engineering tasks with up to 20 hours per task and five trials, Claude Fable 5 leads with 88.2% dominance, ten points above GLM-5.3 (78.1%) and Grok 4.6 (77.9%), according to the BenchLM mirror. BenchLM itself marks it “display only” and does not use it to rank models. Proximal has archived v1, and v2 has 34 tasks, a different harness and a different scale. Benchmarks expire faster than the models they measure, which is what Anthropic was saying this week in other words.

78% stop reading if it smells of AI

Colin Breck cites on his blog a survey by Cynthia Dunlop on how developers react to AI-scented articles: 78% stop reading, 71% avoid the author in the future, and 98% prefer the author’s own writing, flaws and all, to a machine rewrite. Breck describes the rest: design documents generated after the fact for things already built, PR summaries written by machines for machines, and someone who replied to him with an AI summary of his own comments. Bryan Cantrill, at Oxide, requires all public writing to pass Pangram as human-authored. I note it because it affects me. If the reader notices the machine, they leave.

THE DRAMA

Apple removes the “no” switch

David Bushell, a British web developer, recounts on his blog that he turned off Apple Intelligence on macOS 15 in February 2025, when he discovered a feature phoning home to Apple every 15 minutes. Last week he upgraded to macOS 27 and the switch is gone: everything turned back on despite being off, there are Siri processes that cannot be killed, and the package takes up 22.28 GB of disk. With Apple charging £500 per terabyte on new MacBooks, that is £11 of space nobody asked for. The £11 is the least of it. He said no and Apple said yes.

The FBI hack exposes its own hacking unit

The theft of the personal data of at least thousands of FBI employees, including addresses, phone numbers and spouses, includes members of the Remote Operations Unit, the secretive team that builds exploits and tools to break into targets’ devices, according to 404 Media. Almost no public information exists about that unit, not its operations, not its tools, not who is in it. Now a list with their names is in the hands of criminals and, in all likelihood, of some intelligence service.

Kalanick is back and is not apologising

Fortune devotes a long profile to the return of Travis Kalanick, ousted from Uber in 2017 by his own investors, with an industrial robotics company, Atoms, backed by Andreessen Horowitz. “Travis is that guy”, writes Ben Horowitz. Kalanick retracts nothing: “I continue to stand by every decision I made at Uber”, he says, and blames his exit on “extreme wokeness”. A young founder calls him “the Steve Jobs for founders of my generation”. Sarah Lacy, whom Uber threatened to investigate, sums it up from the other side: “the Trump era has given everyone cover that they don’t have to care”. The same people who pushed him out are welcoming him back. Money holds no grudges when there are robots to sell.

Insta360 spends $298 million on chips, more than GoPro is worth

Insta360 has 73% of the worldwide 360 camera market, has just sold its 10 millionth camera, is opening a store in Times Square and has spent $298 million on memory chips, more than GoPro’s entire market value, according to The Verge. GoPro has fallen to 3% of the smart camera market with 200,000 units in the quarter. DJI shipped 4.3 million cameras in a single quarter. Canon, 4.5 million in all of last year. And now Insta360 is preparing mirrorless cameras and glasses. The Chinese killed the action camera, and the DSLR has less time left than it looks.

THE WEEK AHEAD

Sonnet 5.5, Autopilot and the General Court

Anthropic says Sonnet 5.5 and Haiku 5.5 arrive “in the coming weeks”. If Opus’s 40% discount moves down a tier, the price of frontier inference moves again, and that is where I will look first.

Microsoft opens Autopilot in private preview this month and the Home and Code tabs in its Frontier program. I want to see the first usage-based bill of a real customer, because “FinOps for AI” exists for a reason.

In Luxembourg, the General Court has to decide whether to admit the United States into X’s case. It is the first time Washington has entered a DSA fine, and the Commission’s response sets the tone for the ones to come.

Muse Charm ships later this year and Lightspeed starts investing its India fund within two months. And Kevin Roose and Casey Newton launch Machine Gods in 24 days, as Newton recounts. This Friday they go live on Instagram about “the news story we would be talking about if we still had a podcast”. I know which one it is.

REFERENCES

Every source was read directly and is dated between September 18, 2026 and September 25, 2026.

• CNBC: Anthropic, OpenAI hunt for smaller AI data center deals, sources say · September 18, 2026
• laya.convaiinnovations.com: Laya — 33ms Multilingual System 1 Decision Engine with Calibrated Probabilities · September 19, 2026
• blog.colinbreck.com: I Don’t Want to Read What You Didn’t Write · September 20, 2026
• Stratechery: Frontier Overhangs · September 21, 2026
• OpenAI: Introducing GPT-6 Sol and Luna | OpenAI · September 22, 2026
• Anthropic: Introducing Claude Opus 5.5 \ Anthropic · September 22, 2026
• Stratechery: Amazon Blocks Muse, Amazon’s Moat, Aggregator v Aggregator · September 22, 2026
• dbushell.com: I said no and Apple said yes – David Bushell – Web Dev (UK) · September 22, 2026
• Stratechery: More on Muse, Amazon, and Walmart; Muse and Expedia; Whither Google? · September 23, 2026
• 404 Media: FBI Hack Exposed FBI’s Own Hacking Unit · September 23, 2026
• benchlm.ai: FrontierSWE Leaderboard & Scores — September 2026 | BenchLM.ai · September 23, 2026
• Reuters: Cybersecurity startup Island valued at $6.4 billion amid AI agent security risks · September 24, 2026
• TechCrunch: Waymo is scaling fast: Here’s what the fleet data shows · September 24, 2026
• Fortune: Travis Kalanick is back: How Uber’s ousted founder went from pariah to Silicon Valley hero · September 24, 2026
• Platformer: Can Muse make us forget the metaverse? · September 25, 2026
• The Verge: Microsoft thinks its new Copilot ‘super app’ will be as influential as Office · September 25, 2026
• The Register: Uncle Sam wants in on Musk’s €120M fight with Brussels · September 25, 2026
• TechCrunch: Lightspeed targets $250M for new India fund, focusing on early-stage AI · September 25, 2026
• The Verge: Here’s the Tesla Semi… again · September 25, 2026
• The Register: Asahi fork embraces LLMs and lands Linux on the M4 Mac mini · September 25, 2026
• The Verge: Insta360 conquered 360 cameras — now it’s eyeing glasses · September 25, 2026

Thanks for reading.