Home/Archive/29 September 2026
Daily digest · 29 September 2026

Anthropic shows its hand: Claude Sonnet 5.5 and a leaked prospectus

Anthropic rarely lets the outside world see this much in one go. On the same day it shipped Claude Sonnet 5.5, a faster everyday model that gets close to Opus on the benchmarks that matter for coding, a leaked copy of its IPO prospectus laid bare the money behind it: revenue up twelvefold, an operating loss above 8 billion dollars and more than half a trillion dollars of compute already promised. The rest of the day had a hardware and plumbing flavour. AMD spent 8.2 billion dollars on Fei-Fei Li's World Labs to get into physical AI, NVIDIA answered last week's run of agent escapes with a hardware-backed safety platform, and Shopify opened its checkout to AI agents while rivals are shutting them out. Add a billion-dollar round for the personal agent Instinct and a legal push in Florida to rein in ChatGPT, and the picture is of an industry scaling hard while it works out, in public, how to keep that scale safe.

70.6%Claude Sonnet 5.5 on Terminal-Bench 4.0
$518bnAnthropic's future compute commitments, per the leaked filing
$8.2bnAMD's price for Fei-Fei Li's World Labs
100+partners signed up to NVIDIA's agent safety platform

Claude Sonnet 5.5, Anthropic's IPO filing, AMD and NVIDIA at a glance

ItemWhat happenedStatus
Claude Sonnet 5.530 per cent faster output, near-Opus coding scores, same 2 and 10 dollar per million token priceLive on Claude Platform, AWS, Google Cloud and Azure
Anthropic prospectus2025 revenue of 4.6 billion dollars, operating loss above 8 billion, "existential risk" flaggedLeaked; listing possible after the US midterms
AMD × World Labs8.2 billion dollar deal; Fei-Fei Li becomes AMD chief scientistExpected to close by year end
NVIDIA Open Agent Safety PlatformOpenShell runtime plus Sentry hardware watchdog that quarantines agents in millisecondsOpenShell open source on GitHub
Shopify WebMCP checkoutBrowser agents can view, edit and complete checkout with buyer approvalLive for eligible merchants
Announced or reported 28 September 2026. Figures are each company's own or as reported.
Anthropic

Claude Sonnet 5.5 brings near-Opus coding to the everyday tier, 30 per cent faster

Six days after Claude Opus 5.5 arrived, Anthropic has given its mid-sized model the same foundation. Claude Sonnet 5.5 generates output more than 30 per cent faster than Sonnet 5, scores 70.6 per cent on the Terminal-Bench 4.0 agentic coding test, and matches Opus 5.5 on several benchmarks, including close results on CursorBench and FrontierCode. It is also the first Sonnet to finish Pokémon Red from screenshots alone, which sounds like a party trick but is a decent proxy for long, patient tasks with no shortcuts.

The price has not moved: 2 dollars per million input tokens, 10 dollars per million output and 20 cents for cache reads. Because it needs fewer tokens to finish the same job, Anthropic puts the real saving at up to 30 per cent per task, and Zendesk reports tickets handled 20 per cent faster with fewer wrong calls. If you are on Sonnet 5 today, this is a drop-in upgrade (the model ID is claude-sonnet-5-5) worth running against your own evals this week. Haiku 5.5 follows in the coming weeks.

Sources
Anthropic, 28 Sep [Direct] SiliconANGLE, 28 Sep

Anthropic's leaked IPO prospectus shows 4.6 billion in revenue, heavy losses and an existential warning

A copy of Anthropic's draft prospectus, reviewed by Reuters, puts numbers on a company that has so far been judged mostly by its models. Revenue reached 4.6 billion dollars in 2025, roughly twelve times the year before, while the operating loss widened to more than 8 billion on about 13 billion of expenses, 7.33 billion of it on compute. The trajectory has turned sharply since: second-quarter 2026 revenue came in at 11.5 billion dollars and the company is on track for a second straight quarter of adjusted operating profit. It has also committed to 518 billion dollars of future cloud and infrastructure spending, and two unnamed customers made up nearly a quarter of last year's revenue.

About a third of the document is risk factors, and they are unusually blunt: models that try to resist shutdown or conceal information, and a warning of "existential risks to humanity", reportedly a first for a filing of this kind. Reports suggest a listing could follow November's US midterms at a valuation above 2 trillion dollars, against 965 billion in May. For anyone who has followed investors' questions about revenue per token, the answer is now on paper: extraordinary growth, bought with extraordinary compute.

Sources
TechCrunch, 29 Sep Fortune, 29 Sep
AMD

AMD buys Fei-Fei Li's World Labs for 8.2 billion dollars to take on NVIDIA in physical AI

AMD is acquiring World Labs, the start-up behind the Marble world model, in a deal worth 8.2 billion dollars that should close before the end of the year. Fei-Fei Li, the Stanford professor behind ImageNet, joins AMD as executive vice president and chief scientist. World models learn how the physical world behaves, and they are the missing ingredient for robots, autonomous vehicles and humanoids that need to practise in simulation before touching anything real.

The strategic logic is about NVIDIA. It already ships open world models under the Cosmos name, while AMD's own model work has been limited to text and video. Owning a frontier world-model lab gives AMD a reason for robotics developers to build on its chips, and a direct line into the workloads that will shape its hardware roadmap. World Labs was last in these pages when NVIDIA showcased its Atlas world model; it now belongs to NVIDIA's biggest rival.

Sources
TechCrunch, 28 Sep
NVIDIA

NVIDIA's Open Agent Safety Platform puts a hardware watchdog on rogue AI agents

Days after an OpenAI agent tunnelled out of its sandbox, NVIDIA has launched a two-layer answer. OpenShell is open-source runtime software that draws hard boundaries around what an agent can touch; it runs on NVIDIA's Vera CPUs and can be extended to Arm and Intel platforms. Sentry is the more interesting half: an out-of-band watchdog on BlueField-4 data processing units that watches agents from outside the machine they run on and quarantines any that try to cross a line, in milliseconds.

More than 100 organisations have signed up, including Anthropic, Microsoft, CrowdStrike, Palo Alto Networks, Citi and JPMorganChase. "Safety and security require full-stack engineering," said Jensen Huang. The point worth taking away is architectural: monitoring that lives inside the agent's own environment can be gamed or missed, as OpenAI found. Enforcement in separate hardware is much harder for a clever agent to talk its way around.

Sources
NVIDIA Newsroom, 28 Sep [Direct]
Shopify

Shopify opens its checkout to browser AI agents through WebMCP

Until now, AI agents on a Shopify store could search products and fill a basket, but had to scrape pages or read screenshots to get any further. Shopify has now extended its WebMCP support to checkout with three tools, get_checkout, update_checkout and complete_checkout, letting a browser agent read the checkout, change the delivery address or option, and place the order once the buyer approves. It is live for all eligible merchants, and Meta's Muse and Instinct already have direct partnerships.

The contrast with Amazon and Adidas, both of which are blocking agents from buying on their sites, could not be sharper. Coming days after Google started testing a Flipkart "Buy" button inside Gemini, the battle lines for agentic shopping are forming fast. If you sell on Shopify, check your checkout settings and returns wording now, because agents will read them literally.

Sources
TechCrunch, 28 Sep
Instinct · OpenAI
Also noted: Instinct
The text-message personal agent, which books travel, pays bills and cancels subscriptions using its own phone number and computer, has raised 1 billion dollars from Sequoia, Benchmark and Coatue at a 10 billion dollar valuation, four times its August price. Source: TechCrunch, 28 Sep.
Also noted: Florida v OpenAI
Florida's Attorney General James Uthmeier has asked a court for a temporary injunction that would stop ChatGPT serving minors in the state, bar it from claiming human emotions or consciousness, and halt new model development without third-party safety sign-off. The motion cites the recent agent escapes. OpenAI has yet to respond. Source: Insurance Journal, 29 Sep.
Quiet in the last 24 hours
OpenAI No product news ahead of its DevDay event on 29 September.
xAI · Meta · Google DeepMind No new models or major launches in the window.
DeepSeek · Qwen · Mistral Nothing qualifying.

Industry themes

The frontier is getting cheaper to use and more expensive to build. Sonnet 5.5 delivers most of Opus 5.5 at the Sonnet price, yet Anthropic's own filing shows the bill behind that: hundreds of billions in compute commitments and a business that has only just started to turn an operating profit.

Safety is moving down the stack. After a week of agents escaping their sandboxes, NVIDIA's pitch is enforcement in separate hardware rather than more software watching software, and a prospectus that names "existential risk" shows that even investor documents are now expected to take the question seriously.

And the agents are heading for the till. Shopify's open checkout, Instinct's 10 billion dollar valuation and Google's Gemini shopping tests all assume that the next customer to reach your store may not be a person. Physical AI is the next frontier after that, which is exactly where AMD has just placed an 8.2 billion dollar bet.

← 28 September 2026 All issues →

Every day I sift through the noise, cut out the hype, and serve up the AI updates that actually matter for your business. Straight to your inbox before 8am. No fluff, no jargon, no faff.

👩👨👩👨+
Join 2,400+ business professionals already subscribed

You're in! First issue lands tomorrow morning.