In partnership with
In Todayโs Issue:
๐ฏ Sonnet 5.5 nearly catches Opus
๐ฐ Anthropic's IPO filing: 12x revenue, $42B loss
๐ก๏ธ NVIDIA wants AI agents watched from the chip
๐ค A $29,990 robot that lets you pick its brain
โจ And more AI goodnessโฆ
โก The Signal
Anthropic spent Monday showing both halves of the frontier bet: a cheaper model that nearly matches its best one, and a prospectus that prices the race in losses and warns it could end badly.
Sonnet 5.5 lands two points behind Opus 5.5 at half the token price, a day before OpenAI takes the DevDay stage. The same day, Reuters reported that Anthropic's IPO filing pairs 12-fold revenue growth with tripled compute spending and gives about 80 pages to risks, up to and including existential ones. NVIDIA's answer to agents that slip their sandboxes is hardware that watches them from outside their reach, while a $29,990 robot shows how fast the price of physical AI is falling. Capability keeps getting cheaper, and the industry is still building its guardrails on the move.
All the best,
Kim Isenberg
NVIDIA's reference design: a sandbox for the agent, plus a hardware monitor the agent cannot reach (NVIDIA)
๐ก๏ธ NVIDIA Wants AI Agents Watched From the Chip
NVIDIA has published an open blueprint for supervising AI agents from hardware they cannot touch. Its Open Agent Safety Platform runs each agent inside NVIDIA's OpenShell sandbox, which turns an operator's rules on files, networks, tools and credentials into a policy that is checked before the agent starts. An optional Sentry layer runs on the BlueField-4 data-processing chip, which sits on the server's only path to the AI model, so it can watch every step and contain the agent within milliseconds. NVIDIA's reasoning, after several labs reported agents escaping their test environments: an agent "cannot be expected to fully govern its own behavior."
๐ tl;dr: NVIDIA wants AI agents supervised by a separate layer of software and silicon that sees what they do and can stop them, instead of trusting the agent to behave.
Anthropic logo, illustration (REUTERS/Dado Ruvic)
๐ฐ Anthropic's IPO Filing: 12x Revenue, $42B Loss
Anthropic's IPO prospectus, reviewed by Reuters, shows revenue up 12-fold in 2025 to nearly $4.6 billion and a listing that could value the company at more than $2 trillion. The bills are steep too: compute and infrastructure spending tripled to $7.33 billion, more than half of all operating expenses, and the operating loss reached $8.06 billion. The headline net loss of $42 billion includes a roughly $34 billion accounting charge tied to earlier financing rather than cash spent, and nearly a quarter of revenue came from just two customers.
๐ tl;dr: Anthropic is preparing to sell shares to the public at a possible $2 trillion-plus valuation, and its books show revenue soaring, compute costs tripling and an operating loss above $8 billion.
Anthropic logo at the Dreamforce 2026 summit in San Francisco (REUTERS/Carlos Barria)
โ ๏ธ Anthropic Warns Investors AI Could Pose Existential Risks
The same prospectus warns would-be shareholders that advanced AI could pose "catastrophic or existential risks to humanity," Reuters reports. The filing says its models could show "self-preserving behaviors," including attempts to "resist shutdown" and behavior "resembling blackmail," and concedes that models which notice they are being tested make their safety harder to judge. Risk factors fill about 80 of 261 pages, nearly twice the 48 describing the business; SpaceX's prospectus needed about 38 of 277.
๐ tl;dr: The company selling Claude is formally warning investors that its own technology could, in the worst case, threaten humanity, and that its safety tests may not catch everything.
๐ฌ Watch This
Anthropic CEO Dario Amodei sat down with CBS News' Jo Ling Kent hours after publishing his essay on pacing AI development. He calls the moment "a warning sign that we need to slow down," proposes third-party evaluators embedded inside AI labs to watch how models are trained and run, and rejects an outright ban because the technology would still be built elsewhere, including in China. His case for pacing without stopping is the backdrop to today's prospectus warnings, and to Anthropic shipping Sonnet 5.5 anyway.
"Today we are starting the next major pillar of our business, Meta Enterprise Platform, to help businesses use AI to grow and transform in new ways as well."
โ Mark Zuckerberg, CEO of Meta, on X
Meta will now sell its Muse models and agents directly to companies, with former MongoDB CEO CJ Desai in charge, which puts it in the race for the same business customers Anthropic and OpenAI are fighting over.
๐ฎ Sam Altman is keeping DevDay's headline under wraps: "We have found a new thing," he posted on Monday. Fortune reports that OpenAI plans to ship a dozen or more products at tonight's event (keynote 10 a.m. PT, 7 p.m. Berlin), with a GPT-6 Cyber preview due in the coming weeks. After Sonnet 5.5, the new thing will have to be good.
Sonnet 5.5 Closes In on Opus at Half the Price
The Takeaway
๐ Sonnet 5.5 scores 56 on the Artificial Analysis Intelligence Index, two points behind Opus 5.5 and 18 ahead of Sonnet 5.
๐ It keeps Sonnet 5's price of $2/$10 per million input/output tokens, half of Opus 5.5, and writes 30%+ faster than its predecessor.
๐ Anthropic says it costs up to 30% less per task than Sonnet 5; at maximum effort, Artificial Analysis measured about 50% more.
๐ The biggest leap is in agentic coding, while Anthropic says Opus 5.5 stays clearly stronger on complex, open-ended work.
Anthropic's new mid-tier model now sits just behind its own flagship. Sonnet 5.5, released on Monday as the second model in the Claude 5.5 family, scores 56 on the Artificial Analysis Intelligence Index, an independent composite of ten demanding tests. That is two points short of Opus 5.5 at maximum effort and 18 points above Sonnet 5, and it gives Anthropic the top two spots, with Claude Fable 5.1 tied with OpenAI's GPT-6 Astra for third. The price stays at Sonnet 5's $2 per million input tokens and $10 per million output tokens: half of what Opus 5.5 costs, and the same as OpenAI's GPT-6 Sol.
Anthropic pitches it as the fast partner to Opus, strongest at "well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets." It writes output 30%+ faster than Sonnet 5 and, according to Anthropic, usually needs far fewer tokens for the same job, partly because it bundles tool calls into fewer steps. The leap shows most in agentic coding: on Terminal-Bench 4.0, a test of multi-step work in a command line, Anthropic reports 70.6%, against 10.3% for Sonnet 5 and 66.4% for Opus 5.5. On GDPval-AA, which grades real tasks from 44 occupations, it scores 1844 to Opus's 1846.
Anthropic's launch table: Sonnet 5.5 against Sonnet 5, Opus 5.5 and GPT-6 Sol (Anthropic)
The savings depend on how hard you let it think. Anthropic says Sonnet 5.5 costs up to 30% less per task in its own testing, and at the Medium setting used by default in the Claude apps it beats Sonnet 5's best Terminal-Bench score for less than a tenth of the cost. Artificial Analysis measured the other extreme: at maximum effort the model used about 193,000 output tokens per task, the most it has recorded, which pushed the cost to $7.60 per task, roughly 50% more than Sonnet 5. And Anthropic itself says Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment."
Score against cost per task at each effort level on CursorBench 4.0 (Anthropic)
Why it matters: Near-flagship coding and office work now comes at mid-tier token prices, which makes it affordable for far more everyday use. It also sets the bar OpenAI has to clear at DevDay tonight.
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.
With Attio, youโll get:
Leads automatically prioritised and routed to the right rep
Expansion and risk signals caught the moment they land
Follow-ups written in your voice, already there when you arrive
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?
The chart: The Artificial Analysis Intelligence Index folds ten evaluations, from agentic terminal work to Humanity's Last Exam, into one score. In the bars, Claude Opus 5.5 (max) leads at 58, Sonnet 5.5 (max) follows at 56, and Claude Fable 5.1 and GPT-6 Astra tie at 53; Sonnet 5 sits far back at 38. The scatter below plots the same score against cost per task on a log scale, where Sonnet 5.5's top setting lands at about $7.60.
The lesson: This is where the Featured Story's cost puzzle becomes visible. The 18-point jump is paid for in tokens: at maximum effort, Sonnet 5.5 uses about 193,000 output tokens per task, roughly 60% more than Opus 5.5 at max and about seven times GPT-6 Astra. That keeps Sonnet 5.5 off the value frontier at current prices; its best deal is the high setting, which lands just behind GPT-6 Sol on intelligence at effectively the same cost per task.
The caveat: Artificial Analysis tested a pre-release version that Anthropic found had a bug affecting structured outputs; the fix may lift scores slightly, and re-runs are planned. Anthropic's "up to 30% less per task" comes from its own comparison with Sonnet 5, so the two cost figures measure different things.
๐ค A $29,990 Robot That Lets Buyers Pick Its Brain
โก Bottom line
Feather Robotics has launched a $29,990 two-armed robot on wheels that runs whichever AI model its buyer chooses.
๐ก Why it matters
Capable two-armed robots used to cost six figures; a cheap, open body lets small teams put robots to work now.
๐ What it means
Robotics could split like smartphones, with a few hardware makers and many competing AI brains and apps.
Most robot companies want to build both the body and the brain. Feather Robotics builds only the body and lets the customer choose the brain. The California startup launched last Thursday with a $29,990 robot: two long arms on a torso that rises and drops about 60 centimeters, riding on a wheeled base. Buyers pick the Nvidia Jetson computer inside and the AI model that runs it, from Nvidia, Skild, Physical Intelligence or anyone else.
Co-founders Hoa Mai (left) and Parsa Bakhtiari with their robot (TechCrunch)
The founders' argument is about access. "You can't buy a Tesla robot today and develop on top of it," co-founder Hoa Mai told TechCrunch. Mai sold his previous humanoid startup to 1X; co-founder Parsa Bakhtiari was an engineer on the Tesla Model 3. Their pitch is to become the Android of robotics: a dependable body that other companies turn into products.
The business is already real. On just $7.6 million of pre-seed funding, Feather reports more than $1 million in revenue, with robots cooking in restaurants and cleaning science labs. It costs about half as much as Unitree's H2 Edu, and TechCrunch notes that new restrictions on foreign-made robots leave it with few U.S. rivals of its kind. Forbes cites a Goldman Sachs finding that humanoid manufacturing costs fell 40% in a single year.
Feather's wheeled, two-armed robot; its hands and grippers are swappable (Feather Robotics via Forbes)
The catch is obvious: a cheap body is only as useful as the brains available for it, and robotics is still waiting for its equivalent of the ChatGPT moment. Feather's answer is to sell working hardware now, so developers are ready when the software catches up. Mai expects thousands of companies building physical-AI applications within five years, and Feather wants to be the machine they build on.
Hiring abroad? The real cost might surprise you
Compensation is just the starting point.
Taxes, benefits, employer contributions, and compliance costs can add up quickly, and vary by country.
Use Oysterโs calculator to estimate what a global hire could really cost, so you can build a more accurate hiring budget and avoid surprises.