The frontier stretched in two directions at once today, out into the physical world and down in price. Google DeepMind's Gemini Robotics 2 was the headline, giving humanoid robots balance, dexterity and the ability to pass jobs between one another from a single model, while OpenAI reminded everyone that the economics are moving just as fast with an 80 per cent price cut on its GPT-5.6 Luna tier. Underneath sat a quieter but telling pair of releases: Perplexity gave its agents a lasting memory with Projects, and Thinking Machines shrank capable open weights into a model small teams can actually run. Anthropic shipped no product but published a notably candid safety disclosure. The thread tying it together is that the contest is no longer about who holds the single biggest model, but about reach, price and staying power.
Gemini Robotics 2 brings whole-body control and multi-robot teamwork
If you follow embodied AI or robotics at all, this is the standout release of the day. Google DeepMind has shipped Gemini Robotics 2, a family of models that drives a humanoid from its feet to its fingertips rather than just its arms, coordinates several robots at once, and adapts to an unfamiliar machine within a few hours. In demos posted to the company's own X account, an Apptronik Apollo unit handled fine kitchen chores while a mixed team of Apollo and Franka robots tidied a messy garage by handing tasks between them.
The reasoning model, Gemini Robotics ER 2, is already available in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, with the action models open to early-access partners. The practical point for anyone building physical AI is that a single model now spans balance, dexterity and multi-robot handoff, capabilities that until now needed separate stacks. It also resets the bar for every rival chasing general-purpose robot control.
GPT-5.6 gets an 80 per cent price cut on Luna and a faster Sol tier
If you run any automated workflow on OpenAI's API, your bills may be about to drop sharply. OpenAI has cut the price of GPT-5.6 Luna by 80 per cent and GPT-5.6 Terra by 20 per cent, roughly three weeks after the GPT-5.6 family reached a wider audience. Luna now costs 20 cents per million input tokens and 1.20 dollars per million output tokens, down from 1 dollar and 6 dollars, while Terra falls to 2 dollars and 12 dollars per million. The company attributes the savings to efficiency gains found during development, including the model optimising its own production inference code.
Alongside the cut, OpenAI is adding a Fast mode for the top-tier Sol model that runs up to 2.5 times quicker at twice the price, replacing the older Priority Processing option. The signal for builders is clear: the competitive battleground has moved from raw capability to cost per token, so it is worth re-checking which tier your workload actually needs.
Perplexity launches Projects, a persistent workspace built on Computer
This one surfaced on Perplexity's own X channel before any mainstream coverage, so you are reading it early. Perplexity has launched Projects, which it describes as an evolution of Spaces and a hub for ongoing work inside Perplexity Computer. Each Project is a single place to manage, create and collaborate on tasks, with a shared file system and persistent memory that carries context between sessions.
For anyone who has hit the wall of losing state between one-off agent runs, that persistent memory plus a shared file system is the meaningful upgrade, because it turns Perplexity from a question-and-answer tool into something closer to a durable workspace. If you already lean on Spaces for research, expect Projects to become the place where longer, multi-step work lives. It is worth a careful test before you move a real workflow across, since the feature is fresh and detail beyond the announcement is still thin.
Inkling-Small ships as open weights at a quarter of the size
If you fine-tune your own models or care about open weights, Thinking Machines has just handed you a smaller and cheaper option to build on. Inkling-Small is a Mixture-of-Experts model with 276 billion total parameters and 12 billion active, and the full weights are available to download. The company says it matches or beats the larger Inkling on several benchmarks, including instruction following, while delivering more performance per unit of compute.
It keeps the headline features of the bigger model, namely native reasoning over audio and images, adjustable thinking effort, and a context window of up to 1 million tokens. The release was amplified within hours by both NVIDIA AI and Hugging Face, and you can fine-tune it on Tinker or try it in the Tinker Playground. For teams weighing open weights against closed APIs, a model this capable at this size lowers the cost of running your own inference.
Industry themes
The price war is now the main event. OpenAI's 80 per cent cut on Luna shows the fight has moved from who has the smartest model to who can serve intelligence most cheaply, which is good news for anyone paying per token. When a frontier lab trims a working model's price this hard weeks after launch, it is a strong hint about where the margins are heading across the board.
Agents are also growing memories. Perplexity's Projects, with its persistent memory and shared file system, points to a shift from one-off prompts toward durable workspaces that hold context over days rather than minutes. It surfaced through the company's own X channel before mainstream press picked it up, a reminder that official feeds still break product news first.
Embodied and open AI both had strong showings. Google DeepMind's Gemini Robotics 2 pushed whole-body robot control and multi-robot teamwork from a single model, while Thinking Machines put capable open weights within reach of smaller teams. Different ends of the field, but the same message: the next round of competition is being fought on where models can run and what they can physically do, not just on the benchmark charts.