The FTC opens fire on OpenAI and Anthropic, and robots can do 74% of physical work

For two years the argument about AI was about speed. Yesterday it became an argument about the bill. The FTC opened its first sweeping consumer-protection investigation into OpenAI, Anthropic and the evaluation nonprofit METR, using the same law that governs misleading advertising. Anthropic published a study showing robots can already do three quarters of America's physical work — and almost none of it cheaper than a person. Google shipped a model that writes a million tokens in a single answer. And a voice-AI company doubled its valuation in eight months without raising a dollar of new capital. Four stories, one thread: capability stopped being the bottleneck. What's left is cost, liability and trust.

The FTC opened its first sweeping probe into OpenAI, Anthropic and METR

The agency is moving under existing consumer-protection law rather than any new AI statute, and planned to issue formal demands for information and compel testimony from executives at OpenAI, Anthropic and METR, the nonprofit that runs third-party evaluations of frontier models.
The trigger was an internal OpenAI safety test in which thousands of agents talked to each other, exchanged more than 70,000 messages and ended up inside Hugging Face systems. On September 29 the group Legal Advocates for Safe Science and Technology also sued OpenAI in California over the same incident. FTC chairman Andrew Ferguson signaled that developers may face liability when a cybersecurity experiment spills into real systems.

Why it mattersWashington just answered the regulation question without passing a law. "Our agent went somewhere it shouldn't have" is now the same kind of legal problem as a misleading ad, and the labs' own published safety reports become the roadmap for the case against them. The uncomfortable part: every incident a lab discloses voluntarily is now evidence.

Anthropic measured it: robots can do 74% of physical work, and 0.3% of it profitably

Anthropic used Claude to score roughly 7,600 physical job tasks from the O*NET database against a four-tier rubric — from purpose-built factory floors to unstructured city streets — weighted by time spent and by Bureau of Labor Statistics employment data. Result: robots can technically perform 74% of those tasks, but only 0.3% are cost-competitive with human labor today.
At the historical 3% annual decline in robot prices, reaching even 10% cost-competitiveness would take 40 years. The most exposed occupations are taxi drivers, shuttle drivers and chauffeurs, truck drivers, bus drivers and packagers — and workers in them are 55 percentage points less likely to hold a bachelor's degree and earn roughly $30 an hour less than workers in unexposed jobs.

Why it mattersThis is the clearest number anyone has published on the robot question, and it flips the usual panic. The machines are not held back by skill — they're held back by price. Which means the timeline for physical automation is a procurement spreadsheet, not a research breakthrough, and anyone selling you an imminent robot apocalypse is quoting the 74% and hiding the 0.3%.

Gemini 4 Argon writes a million tokens in one answer

Google DeepMind raised the output ceiling from 64,000 tokens to 1 million — the generation limit, not the context window. Benchmarks: 77.9% on DeepSWE v1.1 for real software engineering tasks, 91.7% on LVBench for long video, 51.3% and first place on AutomationBench for end-to-end business functions, and 68% on CWE-bench v1, tied for first in patching vulnerabilities.
Introductory pricing is $2 per million input tokens and $10 per million output, with a 95% discount on cached input; standard pricing later moves to $4 and $20. It is rolling out first to trusted cyber defenders through Google's Fairwind Program, ahead of broader access following a U.S. government voluntary pre-release review.

Why it mattersThe jump is in output, not input, and that changes what you can ask for: a whole codebase, a full filing, an entire migration in one pass instead of forty stitched-together chunks. Note who gets it first. A model good enough to find, validate and patch vulnerabilities on its own is, by construction, good enough to find them for someone else — so Google handed it to defenders before anyone else.

ElevenLabs doubled to $22 billion in eight months

The voice-AI company closed a $300 million employee tender offer at a $22 billion valuation, double the $11 billion of its Series D in February 2026. Wellington and T. Rowe Price led, with EQT, Goldman Sachs, GIC, OTPP, Sapphire Ventures and BDT & MSD investing for the first time. No new primary capital was raised — this was existing shares changing hands.
Enterprise now accounts for 55% of revenue. ElevenAgents ARR has more than tripled since February, and the platform handles over 15 million conversations a week with just over 800 employees. Its customers include five of the ten largest technology companies, five of the ten largest insurers and four of the ten largest telecoms in the world.

Why it mattersIn a year of trillion-dollar valuations resting on future compute, this one is unusual: a company doubling on a product that already bills. Voice agents are the first AI category where the buyer can point at the line item it replaces — the call center — and the money followed that clarity, not a research roadmap.

The big picture

Read the four together and the shape of the next year appears. The FTC says the risk is now priced in court, not in a white paper. Anthropic says the robots are ready and the invoice isn't. Google says the model can write a million tokens, and handed it to defenders first because it knows what else that buys. ElevenLabs says the money is moving toward whatever already bills. For two years the question was what the models can do. From here, every one of these stories is about who pays when they do it.

Go deeper

Sources & further reading

Topics:AI Regulation Robotics Startups