OpenAI rolls out Computer Use tools for ChatGPT browser control

OpenAI brings Computer Use to ChatGPT and Codex

OpenAI is rolling out Computer Use tools that allow ChatGPT to operate a computer and web browser. The company launched a Chrome extension that lets ChatGPT take over the browser, created cloud and in-app browsers for ChatGPT to interact with public websites, and added Computer Use to its Codex coding tool. The same underlying technology lets users have ChatGPT complete tasks on other apps. OpenAI president Greg Brockman wrote on X that an April update made the technology “no longer just for coders, but for anyone who does computer work.” OpenAI staffers say the tools are at an inflection point. Rival Anthropic was first to market with its version of Computer Use in 2024; as of May, at least 600,000 organizations had tried Anthropic’s Claude Cowork feature. OpenAI says its technology’s capabilities have caught up.

Ari Weinstein, a manager on OpenAI’s Computer Use team, said, “Once ChatGPT can use computers and software faster than you or I can, it’s going to change the way that you, by default, want to interact with your computer.” Training for Computer Use involved several phases, including frame-by-frame screenshots and datasets where people demonstrated how to complete tasks. OpenAI also likely used reinforcement learning, where the model is rewarded for successful behavior. Weinstein said OpenAI upgraded how its models interact with a computer: previously, ChatGPT would take a screenshot, analyze pixels, inject a command, and repeat; now, with a website open, it can read the page’s memory structure, accessibility information, and link connections. The tool still takes screenshots; users are asked to let ChatGPT record their screen during setup. Cristian Medina Ruiz, a hobbyist coder in the Czech Republic, used Codex to rebuild the 2013 city-building game “SimCity” from its original source code. Weinstein and James Sun, who works on browser capabilities at OpenAI, said part of the technology’s promise is connecting ChatGPT’s agents to the “long tail” of websites, automating data entry work, compliance tasks, and calendar scheduling. Sun said his team has been grappling with “confirmation policy,” deciding that whenever an AI agent transmits data or deletes something, it should ask the user first. “You can make something 100% safe, but then it’s just very, very difficult to use, and you can make something that is very easy to use but very dangerous,” Sun said. “So we’re really trying to figure out how to build something that’s beneficial and also easy to use.”

Inside OpenAI’s plan to make ChatGPT the way you use your computer →

China’s AI lead narrows as Western models keep edge in cyber and reliability

Chinese open-weights models from Moonshot, Alibaba, and Z.ai now sit near the top of almost every broad, demanding evaluation, according to a Frontier Radar report. The often-cited gap of a few months between Chinese and Western models has shrunk enough to become an investor problem. Anthropic is fielding uncomfortable questions ahead of its upcoming IPO and points to its remaining lead at the top in its defense. Below that tier, the field belongs largely to open, far cheaper models from China. Two accusations are in play: Chinese labs allegedly tapped Western models as teachers through distillation, and allegedly tune their models for strong benchmark scores without broad capabilities to match, a practice called benchmaxxing. At launch, Artificial Analysis had K3 in third place on its Intelligence Index with 57 points, behind GPT-5.5 and Opus 4.8. On AutomationBench-AA, K3 debuted in first place until Anthropic answered with Opus 5. On CEO-Bench, where an agent runs a fictional software company for 500 simulated days, K3 posted the best published single run at $22.15 million. Qwen3.8-Max reaches a similarly high overall level.

Three areas of Western lead remain measurable. The first is abstract specialty tests: Opus 5 retook the top of the index with 61 points, a small gap over K3’s 57; on ARC-AGI-2, the gap widens to 60.4 versus 89.2 percent. The second is reliability: the AA-AnalystAgent benchmark counts a task as solved only if a model gets it right in five out of five independent runs, and Opus 5 leads with 54 percent, ahead of GPT-5.5 at 50, while K3 is the best open model at 39 percent. The third is cybersecurity: a joint assessment by the UK’s AISI and the US CAISI found that K3 lags far behind leading US models in offensive cyber capabilities, scoring 32 percent on ExploitBench versus about 76 for top US models, and failing all 41 tasks that required executing code on a target system. GLM-5.3, unveiled August 14, scores 54.4 percent on ExploitBench by Z.ai’s own measurement, more than double its predecessor GLM-5.2, and edges past leading US models on finding and validating vulnerabilities in source code. Anthropic’s best cyber model, Mythos 5, hits 78 percent on ExploitBench but is only available under controlled conditions through Project Glasswing. The White House has declared distillation campaigns against US models a national security threat by memorandum, and Anthropic’s Mythos 5 and Fable 5 are now under export controls, the first time the US government has controlled access to a model itself.

Frontier Radar #4: China has caught up, so what’s left of the Western AI lead? →

AI pioneer Sutton calls synthetic data a big mistake

AI researcher and Turing Award winner Richard Sutton, a founder of reinforcement learning and author of the 2019 essay “The Bitter Lesson,” has called synthetic data a “big mistake” for solving the scaling problem of large language models. In a recent conversation, he introduced his new company, Oak Lab, which he founded with his former student Khurram Javeed. Sutton said large language models are both a good and bad example of his “Bitter Lesson” argument that AI methods scaling with compute win out. They scaled enormously with compute and could “drink in the internet,” but hit a wall because the internet is finite, while the real world is “massively bigger than everything we stored on the internet.” He said leaning too heavily on human knowledge ultimately holds the systems back.

Asked whether synthetic data could break through this bottleneck, Sutton said: “No, that’s just a big mistake.” The reasoning comes from the “Big World Hypothesis” formulated by Javeed, which holds that the world is infinitely complex and “massively more complex than your mind than any agent.” Any simulation of it is “microscopic” and fails to match reality, with wrong friction values or inaccurate models of a robot’s motor behavior. Sutton also said the world contains many other agents whose inner workings cannot be generated as synthetic data: “There’s no way we can have synthetic data for other people’s minds.” A second objection is the human bottleneck: by Javeed’s argument, deciding which synthetic data is good or bad requires human experts. “You need human experts who know what’s a good data set and what’s a bad data set for that approach to scale. So it is bottlenecked by humans.” Sutton’s alternative is to take humans out of the loop and let agents learn from their own experience, learning their own world model and correcting it instead of relying on a frozen simulation model humans built: “Simulators they make themselves.” He criticized today’s language models for stopping learning after training, saying “their weights never change.” He called for real continual learning, since “all learning is continual,” without wiping out old knowledge, a problem known as catastrophic forgetting. Sutton said this problem can be solved in part with a method called “Continual Backprop” that his team published in Nature. He called language models an “amazing scientific breakthrough,” but only “like 20% or a quarter of intelligence.”

KI-Pioneer Sutton calls synthetic data a “big mistake” in the face of an infinitely complex world →

Anthropic runs unreleased Model 2 internally, outperforming Mythos 5

Anthropic is running an unreleased AI model internally that outperforms every publicly available version of Claude. The company’s Risk Report from August 2026 calls the model “Model 2” and places it in the Mythos class. Anthropic says it is slightly stronger overall than Claude Mythos 5, but weaker in some areas, and does not show a capability jump as large as the one from Opus 4.6 to Mythos. On the company’s internal capability index, AECI, it sits about 1.5 points above Mythos 5, a gain smaller than the jump from Mythos Preview to Mythos 5.

Internally, the company uses the model heavily for coding, data generation, and research and engineering, sometimes through agents that run continuously. Claude now writes most of the code in Anthropic’s production systems. Model 2 went through an internal review before deployment but was not tested as thoroughly as Mythos 5. Anthropic found no new or more worrying misalignments in the process and rates the overall risk from misalignment as “low.” There are no plans to release the model externally right now.

Anthropic’s most capable model, codenamed “Model 2,” is for internal use only →

Unitree surges 460% in Shanghai debut on circular AI financing

Unitree Robotics’s stock jumped as much as 629 percent during its Shanghai debut, closing up 460 percent, according to Bloomberg. The Hangzhou company raised 6.1 billion yuan ($904 million), becoming the first publicly traded maker of humanoid robots on the Chinese mainland. The closing price valued Unitree at around $50 billion. Behind the surge is a circular business model, according to a Financial Times report. Chinese humanoid makers sell a large share of their machines to state-backed training centers, where people “teach” robots physical tasks through teleoperation. The centers then sell the collected data back to the manufacturers. Local governments and manufacturers often fund the centers together.

“It also blurs the distinction between independent demand and demand created within a policy-supported ecosystem,” said analyst Poe Zhao. By June there were more than 90 such centers, according to Interact Analysis. Training data for a five-minute robot dance can cost up to one million yuan ($148,000). At manufacturer Leju, training centers accounted for 45 percent of sales of its flagship robot. At Unitree, nearly three-quarters of humanoid revenue in the first nine months of 2025 came from education and research. “Both sides get what they need, and both sides have something to show,” one algorithms engineer said. Marco Wang of Interact Analysis said the data isn’t “100 percent useful,” since the robots don’t run in real-world settings. Only two to three of every eight training hours are usable, one center manager said. Unitree was valued at 35.89 times revenue, according to Bloomberg, compared with about 20 times for its Hong Kong rivals. “There’s clearly no fundamental basis for the share price surge,” said Vey-Sern Ling of Union Bancaire Privée. Early-stage investors are already looking for an exit, the FT reports. About 20 percent of the IPO went to strategic investors, including AI startup Deepseek. Supporters point to electric cars and solar panels, which China now dominates after state-driven demand got them started.

China now has its own AI circular financing scheme →