Google is cutting free Gemini access. Starting October 9, accounts without a paid subscription are limited to Flash-Lite, Google's lightest model, and even the AI Plus plan loses access to Pro, according to Les Numériques. The Decoder reports a larger cut is already scheduled: free Gemini access ends entirely in October 2026, leaving non-subscribers on the smallest model while Flash and Pro stay behind a paywall.
The same week produced a cluster of open-weight and self-hostable alternatives from NASA, IBM, DeepSeek, Alibaba and Aleph Alpha. None of them replace Gemini. But together they sharpen a question that matters more than most spec sheets when buying AI-driven hardware: does the intelligence inside a product live in a vendor's cloud, where it can be downgraded or re-priced on that vendor's schedule, or can it run on a model the buyer or manufacturer controls?
What shipped the same week
NASA and IBM released the Lunar Foundation Model, which The Decoder describes as one of the first open-source AI models built for lunar science, trained on nearly 2 million tile bundles drawn from 17 years of Lunar Reconnaissance Orbiter data. It is a research tool, not a robotics product, but it shows a major institutional AI effort choosing open weights over a closed API.
MarkTechPost reported several releases pointing the same direction:
- DeepSeek shipped official macOS and Windows desktop apps for DeepSeek Harness v0.2, its MIT-licensed open-source agent harness, adding a plugin manager and scheduled automation tasks.
- IBM made self-hosted deployment of its agentic software development platform, IBM Bob, generally available for on-premises, private, sovereign cloud and air-gapped environments.
- Prime Intellect launched Prime Inference, an OpenAI-compatible serverless and reserved serving platform for frontier open models.
- Alibaba's Qwen, an invite-only chatbot in April 2023, had grown into a 2.4-trillion-parameter open-weight model by August 2026.
- Aleph Alpha released Kolibri, a 78.1-billion-parameter open-weight Mixture-of-Experts model under an Apache 2.0 license.
None of this means open models beat the best closed ones. It means the open-weight tier now includes large, serious options that can run on infrastructure a buyer or manufacturer controls, instead of infrastructure the model vendor controls.
Why this matters for a robot, not just a chatbot
A humanoid or companion robot's "intelligence" is rarely self-contained in the hardware. For many products on the market, perception, language understanding and task planning run at least partly on a cloud model the manufacturer licenses from a provider like Google, OpenAI or Anthropic. That arrangement is invisible during a demo and in the first weeks of ownership. It becomes visible the moment the provider changes terms.
That is what just happened to Gemini users: a capability that was free or included in a mid-tier plan on October 8 requires a different, pricier subscription on October 9. The hardware did not change. The economics of using it did. A robot wired to a single closed API inherits that risk: the manufacturer's cost base can shift overnight, and the cost tends to land on the subscription price, the feature set, or both.
This is not an argument that open models are automatically better for a robot. A vendor-managed closed API can mean better support and more consistent updates. The tradeoff is control versus convenience, and convenience has a price the vendor sets unilaterally, not the buyer.
What to ask before buying
Four questions separate a platform with real optionality from a thin shell around someone else's API:
- Which model runs the core functions? Ask whether navigation, voice interaction and task planning depend on a named closed model, or can run on an open-weight model.
- Can it run offline or self-hosted? IBM's move to make Bob available in air-gapped deployments shows self-hosting is a shipping option for agentic systems now, not a research goal. Ask whether the robot's vendor offers anything comparable.
- What happens if the cloud tier changes? Gemini's October 9 cut previews a risk that applies to any subscription-gated AI feature. Ask what the robot does, and at what cost, if its model's pricing or access tier changes.
- Is there a fallback mode? A robot that degrades gracefully to local or open-model processing when the cloud connection or subscription lapses is a different purchase from one that stops working.
There is no universally correct answer. A buyer who wants the most capable assistant today and accepts subscription risk can reasonably choose a closed-API product. A buyer prioritizing long-term cost control or offline operation should weight self-hosting and open-weight compatibility more heavily. Either way, ask before buying, not after. Compare power, payload and battery life on the humanoid comparison tool and track pricing over time on the humanoid robot price trackerbut treat the model behind the hardware, and who controls access to it, as a spec in its own right. See our AI news coverage for more on this shift and our humanoid robot rankings for how it factors into current scores.