Autonomous Figure AI Robot Upgrade, WHRG 2026, Free Ox Alpha Stealth

Is the humanoid robot race going to be won by numbers or intelligence? Today on AI News, we're ranking the newest humanoid robot feats for you to decide who's leading. Plus, we'll look at a totally free multimodal mystery model. So like and subscribe, and let's start with humanoid robot feat number one. This is figure 03, learning to autonomously climb a ladder. And when CEO Brett Adcock was asked if it would be able to handle two times the speed, he said it would, but for now, they're starting out going slow, and in time it will get much faster. And what you're seeing here is a combination of the Helix 02 system. This is a full body loco manipulation system with three parts. With system two, the model is using a slow reasoning layer here to break up the instructions like approach the ladder, establish a grip, shift the weight, advance the limb, and repeat. And then it hands off semantic goals to the layers below rather than a planned individual joint movement. So this then goes to system one, which turns that goal into full body joint targets at 200 hertz, and it pulls from every sensor at once. In fact, the head cameras track rung spacing and ladder angles, and then the palm cameras mounted in the hands feed visual data even when the head camera's view is blocked by the robot's own arms during a reach. And actually, when climbing back down, this system reverses the sequence, which is an even harder balance problem to solve. System two sequences the descent. It locates the next rung down without a direct line of sight, and since the head camera faces mostly forward and down, not straight beneath the robot, the palm cameras are doing more of the work here, effectively acting as the eyes of the robot. Then system one shifts from mostly visual guidance climbing up to more tactile guidance climbing back down, leaning on fingertip sensors to confirm the grip, and updating at one kilohertz. And something to note here is that descending also means lowering the body's weight against gravity rather than climbing against it, so system zero's balance correction is managing this direct force profile at every step. And so far, the same helix zero two system has demonstrated quite a few other tasks, including washing dishes and cleaning rooms and more, as well as climbing stairs. But in another embodied AI demo, we have feat number two. This is Galbot, and it's showing its Astra tennis, where this robot was able to autonomously volley over 130 tennis returns without error. And it also performed with humans in the mix. This was at the World Humanoid Robot Games in China 2026, where Galbot performed these tennis feats. And the third feat was accomplished by Tiangong at the World Humanoid Robot Games of 2026. Here, their humanoid was able to improve on the 1500 meter world record by one minute and six seconds, completing the entire run in two minutes and 21 seconds. Then, Tiangong also used its omni robot to take the gold in the 400 meter run in just 45.66 seconds. Additionally, another Tiangong robot also won the 400 meter final by finishing in just 38.15 seconds, breaking another human world record once again. And finally, they were able to complete a 100 meter run in just 9.32 seconds, breaking Usain Bolt's human world record. And the fourth feat is from Booster Robotics. Here at the World Humanoid Robot Games 2026, they were able to synchronize a performance using 80 separate T2 humanoids. And they also carried out a rehearsal with these robots for the opening ceremony of the World Humanoid Robot Games 2026. But tell us down in the comments below if you know what this device is on the head of the robot. Is this a 360 LiDAR system or what? And finally, in the latest in generative AI, a new AI model called OX Alpha appeared on the developer platforms OpenRouter and OpenCode on August 20th, with no company name attached to it. And OXAlpha.com has since emerged as a free no signup browser interface for chatting with it directly. In fact, OX Alpha is listed under the provider name Stealth, and the model ID Stealth slash OX Alpha build as a reasoning model built for coding, sustained agentic work and production tasks. And it carries a 1 million token context window with a maximum output of 131,000 tokens, accepting text, image and video inputs. Independent Needle in a haystack testing also found the context window functions close to that stated size, with retrieval succeeding at placements around 934,000 tokens and failing above roughly 1.005 million tokens. And pricing during this preview period is $0 per million tokens for both input and output, with OpenCode's own announcement stating the model would remain free for about a week from its launch date, putting the likely cutoff around August 27th, though OpenRouter's own listing doesn't show a hard end date. And claimed daily processing capacity varies by source, with OpenCode citing 100 trillion tokens per day, with another provider citing 1 quadrillion. And when it comes to adoption, within roughly three days of the launch, OpenCode's own usage data showed close to 16 trillion tokens processed, 221,000 unique users, and just over 5 million sessions, making it the platform's second most used model by recent activity. Then, on benchmarks, early claims were more impressive than what the fuller test later showed. For instance, a widely shared figure of an 80% score on Deep Software Engineering coding benchmark came from a hand-picked set of just 10 tasks. Then, a more complete run across all 113 Deep Software Engineering tasks put the resolved rate at 58.4% instead, which was a meaningfully more modest number than the one that initially circulated. But data handling policies differ depending on which platform a person accesses the model through, and the two publicly stated policies aren't identical. OpenRouter's stealth model terms permit the model's provider to use submitted content for training, evaluation, and improvement. In contrast, OpenCode's own listing states zero data retention and no training use on its specific route. So far, no organization has claimed responsibility for building OX-Alpha, and community investigation has focused on technical fingerprinting rather than an official statement. Plus, several of OX-Alpha's error message formats reportedly match those returned by Jupu's GLM-branded models. But OX-Alpha's origin is still unconfirmed, with other sources claiming that this is Gemini's newest model. Anyways, like and subscribe, and tell us in the comments below where you think OX-Alpha is from, and thanks for watching!