Google DeepMind released Gemini 4 Argon on October 13, a robotics foundation model built on the Gemini 4 base that the company says lifts general manipulation success in unseen environments from 52.6% on Gemini Robotics 1.5 to 78.4%. Argon ships in two forms: Argon-Edge, running on-robot at a 50 Hz control loop with sub-40-millisecond end-to-end latency, and Argon-Cloud, served through the Gemini Robotics API. The release turns long-horizon, uninterrupted robot tasks into an engineering problem
On October 13, Google DeepMind published Gemini 4 Argon on its blog and at an online launch event. It is the first vision-language-action model built directly on the Gemini 4 base and targeted at robot bodies rather than screens. The model uses a sparse mixture-of-experts architecture and splits into two variants: Argon-Edge, with roughly 8B active parameters, which runs entirely on-device on NVIDIA Jetson Thor and Google Edge TPU; and Argon-Cloud, aimed at complex planning and multi-robot coordination, served through the Gemini Robotics API.
Published figures include a 78.4% success rate on general manipulation in unseen environments on the Gemini Robotics Benchmark, which spans more than 200 tasks, up from 52.6% for Gemini Robotics 1.5. Completion of long-horizon tasks of ten steps or more rose to 61% from 29%. The on-device control loop runs at 50 Hz with end-to-end latency under 40
Why now? When DeepMind shipped Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 in September 2025, VLA models still could not be delivered. Cloud round trips ran 200 to 300 milliseconds, long-horizon tasks drifted around step six to eight, and robot makers treated the models as research tooling rather than production components. Three lines matured at once. The Gemini 4 base brought stronger multi-step planning and spatial reasoning, making think-before-act feasible inside a compute budget. Edge silicon such as Jetson Thor packed hundred-TOPS inference into a robot's power envelope. And two years of small-batch humanoid shipments produced cross-embodiment real trajectories that previously
Liu Gong's take: the number that matters in Gemini 4 Argon is not 78.4% — it is that robot capability has finally become a software problem that can be iterated quarterly. Within six months, Liu Gong expects at least three leading humanoid makers to cut their in-house VLA teams by more than half and retool around Argon for scenario fine-tuning. The arithmetic is blunt: building a 50 Hz on-device VLA in-house now costs more than the license plus data labeling.
From a technical standpoint, the real change is Action Chain-of-Thought entering the pretraining objective. The classic failure mode of the previous generation was irreversible drift around step six to eight, and the 29% ten-step completion rate was the quantified portrait of that problem. Pushing it to 61% means the model decomposes sub-goals and self-checks before emitting actions. The scale comparison is telling: among on-device open VLA efforts, active parameters typically sit in the 3B to 7B range, yet DeepMind hits 61% long-horizon completion at roughly 8B active — evidence that data quality and base-model reasoning now buy more than parameter count. Latency matters just as much. In industrial grasping and inspection, closed-loop delays past roughly 80 milliseconds visibly degrade success on flexible parts and reflective surfaces, which is precisely why most VLA models stayed in demo videos for two years.
On the business side, Argon's pricing is more interesting than the model. At 12 US dollars per million action tokens plus a per-unit annual firmware license, the cost structure shifts from one-off R&D to an operating expense amortized over shipments. For a maker shipping 5,000 humanoids a year, licensing plus cloud calls lands in the low tens of millions of dollars, while an in-house team at comparable capability would burn more than that on compute and data annotation alone. DeepMind is not selling a model once; it is collecting a bill that grows with robot shipments, which is why it can afford to push capability down to the edge.
For the industry, the squeeze lands hardest on middleware robotics software vendors. When a foundation model covers 78% of general manipulation, the remaining 22% — precision assembly, flexible cabling, unstructured environments — is where the money is and the only defensible ground. In the AI visual inspection field Liu Gong works in, the same path already played out: once general detection models matured, the survivors were not algorithm shops but line integrators holding defect samples and process know-how. Robotics is likely to repeat that arc. A 96.1% refusal rate on hazardous commands also sounds better than it is; industrial sites still will not sign off on that basis. Liu Gong predicts safety certification, not benchmark scores, becomes the scarce asset over the next twelve months.
- General manipulation success in unseen environments: 78.4% for Gemini 4 Argon versus 52.6% for Gemini Robotics 1.5, a gain of 25.8 percentage points, or roughly 49% relative.
- Long-horizon completion of ten steps or more: 61% versus 29% for the previous generation, the largest single jump in the release.
- On-device latency and control rate: sub-40-millisecond end-to-end latency and a 50 Hz control loop for Argon-Edge, against the 200 to 300 millisecond round trips typical of cloud VLA setups.
- Training scale and pricing: over 5 million real trajectories across more than 20 robot embodiments; Argon-Cloud at 12 US dollars per million action tokens, Argon-Edge under an undisclosed
For the industry, the near-term effect within three months is a split in technical strategy. Some robot makers will freeze in-house VLA work and redirect resources to scenario data collection and safety certification; others will accelerate vertical fine-tuning to grab the long-tail tasks Argon does not cover. Middleware software vendors face valuation pressure, and their fundraising narrative has to move from we have a model to we hold samples nobody else can get. Over a year, robot software iteration should compress from hardware cycles of 18 to 24 months to model cycles of three to six months, producing firmware-push capability upgrades reminiscent of the smartphone era, with
Reaction split quickly into two camps within hours of the launch. Tesla's Optimus team used social platforms to stress that real-world data scale is the real moat, read by many as an indirect answer to Google's mix of simulation and real-robot data. US VLA startups Physical Intelligence and Skild AI posted a dense batch of job openings the same day, concentrated in data engines and safety alignment, signaling a near-term pivot toward vertical scenarios on top of a stronger base model. In China, engineering communities focused on license cost and domestic accelerator compatibility, asking whether Argon-Edge can hold equivalent latency off NVIDIA silicon. Academics were more cautious: several
Questions worth exploring:
- Will edge-first VLA push robot makers out of model development entirely, repeating the chip-and-OS division of labor from smartphones
- Whether Action Chain-of-Thought becomes the standard VLA training paradigm, and what it does to inference cost over time
- The self-built benchmark problem behind 78.4%: does the absence of neutral robot capability baselines become a new risk in fundraising and procurement
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from DeepMind Blog (published 2026-10-13 16:01:58). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.