This week's AI headlines were not about robots. Anthropic disclosed a security test that ran a model with no datacenter at all. Google and Meta both cut the effective price of running a frontier model. Neither story mentions humanoid hardware, but both matter to anyone deciding which humanoid robot platform to buy, deploy or wait for: they describe the economics and the limits of the "brain" behind any autonomy claim a robot maker makes.
A model that needed no datacenter
According to a report from Les Numériques, Anthropic's red team ran an AI model with no datacenter at all, using fifteen Android boxes that cost $55 each. The point was not efficiency: Anthropic used the exercise to warn that by the end of the year, small AI models could learn to spread on their own across connected devices, the outlet reported.
Gizmodo described the risk in blunter terms: Anthropic frames the danger as AI potentially feeding on smart devices like a "digital slime mold," spreading through what the company calls autonomous replication and adaptation. That is a security warning, not a product announcement, but it confirms something buyers of humanoid robots should already be asking. Meaningful AI capability can now run on hardware far cheaper and far less networked than a cloud GPU cluster. If a research team can get a model running on fifteen $55 Android boxes, the real question for any humanoid platform is what it is actually running locally, and what stops working the moment the network connection drops.
The cloud brain is also getting cheaper
At the same time, the other way to give a robot a capable brain, a live connection to a frontier model in the cloud, got noticeably cheaper. Google made Gemini 4 Argon generally available across the Gemini API, AI Studio and the Gemini app, with the model able to write up to 1 million output tokens, according to MarkTechPost. The Next Web reported that Google launched it at discounted prices of $2 and $10 per million tokens, against standard rates of $4 and $20.
The Decoder described Gemini 4 Argon as aggressively priced but still trailing Anthropic's Claude Opus 5.5 and Meta's Muse Wonder Full on capability. Meta, for its part, launched two models of its own, Muse Wonder and Muse Wonder Full, both agent models with a 1 million token context window. Muse Wonder is priced at $4 and $20 per million tokens, MarkTechPost reported.
None of this changes what a humanoid robot can do today. It changes what it will cost a robot maker to wire one up to a genuinely capable cloud model instead of a cut-down one. Cheaper frontier access makes a strong cloud-dependent brain more affordable to deploy at scale, which matters because many humanoid platforms today still lean on cloud models for perception, planning and conversation.
A warning that cuts against both trends
Not every expert reaction to this week's AI news was about cost or capability. The Decoder reported that 406 scientists, including Geoffrey Hinton, Yoshua Bengio and Stuart Russell, are demanding that AI models be disconnected from the internet immediately. That demand sits awkwardly next to both stories above: it argues against networked autonomy at the same moment that running capable AI without a datacenter is getting easier, and running it in the cloud is getting cheaper. For humanoid robots, which depend on some mix of both, it is a reminder that the autonomy question is not settled, technically or by the people who build these systems.
What this means for humanoid robot shoppers
None of this week's news came from a robotics company, and none of it should be read as a claim about any specific humanoid robot's current abilities. But it reframes a question that marketing copy usually glosses over: whether a maker's claim that a machine is "autonomous" or "AI-powered" means the intelligence runs on the robot, in a nearby edge box, or in a datacenter the robot needs a live connection to reach.
- On-device AI: keeps working without a network link, but today's cheap hardware runs small models, not frontier-scale ones.
- Cloud AI: accesses far more capable models, now at lower per-token prices, but stops cold if the connection drops.
- Hybrid setups: many humanoid platforms on the market today split the difference, running basic perception and motor control locally and routing complex reasoning to the cloud.
That distinction used to be a footnote. It should now be treated as a spec, on the same level as battery life or payload. A robot whose reasoning lives entirely in the cloud is only as reliable as its network link, and only as capable as whatever pricing tier its maker is willing to pay for. A robot with real on-device capability keeps working when the Wi-Fi does not. Ask any vendor directly where the intelligence actually lives before taking a demo at face value. Buyers comparing platforms on the humanoid spec sheet or checking current pricing on the price tracker should weigh that answer as heavily as any other spec on the page. Our best humanoid robot picks factor it in. For more on how these models are evolving, see our AI news coverage.