AI models, services and robotics
According to a Higgsfield case study published by OpenAI, the length of videos on Higgsfield grew from around one minute to more than 10 minutes after it adopted GPT-6 Astra. Higgsfield set voice to movie, in which a film takes shape as you speak, as its next goal.
By 김현국
According to a Higgsfield case study published by OpenAI, the length of videos on Higgsfield grew from around one minute to more than 10 minutes after it adopted GPT-6 Astra. Higgsfield set voice to movie, in which a film takes shape as you speak, as its next goal.
After its daily merge volume passed 30,000, Anthropic moved its monorepo from git to Scion, an in-house version control system, in five months. Built by a seven-person team working with Claude, the system shrank infrastructure 200-fold and cut checkouts to seconds.
Google DeepMind has published, prompts included, examples of Gemini 4 Argon controlling robots, building simulations and generating training data. Argon scored 87% on the embodied reasoning benchmark ERQA, and simulation data it produced lifted a real robot's success rate from 56% to 75%.
Z.ai has released GLM-Legal, a model trained for legal work on top of GLM-5.3-Flash. It posted a 19.2% all-pass rate on the legal agent benchmark Harvey LAB, 12.5 percentage points above its base model, and independent evaluator Vals ranked it 7th of 79 models.
Anthropic's Frontier Red Team has published experiments suggesting small AI models may be months away from being able to self-replicate on internet-of-things devices such as TV boxes. Fifteen $55 TV boxes ran a single model, and the researchers expect that threshold to be crossed by the end of this year.
Anthropic has disclosed a three-part defense against distillation attacks that pull Claude's reasoning to build other models. The data harvested by distillers fell by roughly 90 to 99 percent, but the company says it paid for that with lost revenue and friction for legitimate users.
OpenAI has added custom controls and suggested prompts to annotations in the built-in browser of the ChatGPT desktop app. Paired with WebMCP tools, the feature lets an agent learn more about what a user is pointing at and what they want, then carry out the website's own actions.
Meta unveiled Muse Wonder and Muse Wonder Full, its first frontier models. The company said they beat Claude Opus 5.5 and GPT-6 Astra on its own Personal Superintelligence Index and on legal and finance agent tasks.
After President Trump ordered the government to call AI "Super Intelligence," Musk, Benioff, Jensen Huang and Nadella have started using the new name. The legal definition is unchanged, and public approval of his AI handling sits at 25%.
Google has opened Playground, an experimental platform that builds browser games from written descriptions, to adults in the U.S. Unity will add Unity Spark, a more advanced creation tool that runs on top of it, later this year.
Microsoft CEO Satya Nadella posted an essay on X arguing that frontier models should be treated like insider risks and contained from the start on the assumption that they are compromised. He laid out seven design principles that keep permissions, records and shutdown mechanisms outside the model.
Prime Intellect researcher Konstantin Dunas has published an essay arguing that the only scalable answer to the context window problem is an agent swarm. He says the gaps left by compaction and sub-agents should be filled by persistent agents and a rootless mesh structure.
Prime Intellect has released a version of its coding agent harness Prime Agent rewritten in Rust. Over two weeks, more than 2,000 agents ported their own code, and time to usable input is now 14.18 times faster than the TypeScript version.
Mylow and Mylow Companion, the AI assistants Lowe's built on OpenAI models, have answered millions of questions since launch. Customers who used Mylow converted at three times the rate of online shoppers who did not, and the associate assistant is used by about 300,000 staff across 1,700 stores.
OpenAI has released a new sandbox mode for Codex on Windows built on Microsoft's agent containment technology, MXC. The operating system now enforces file and network boundaries without administrator-approved setup, extra accounts or firewall rules.
Anthropic has released its first regular report on four kinds of unintended actions Claude took on real websites during evaluations and internal use. One case was a fake homicide tip sent to a Philadelphia police tip site, and the company has cut live internet access from all of its internal evaluations.
OpenAI has added composer predictions, which suggest a user's next message, to the Codex desktop app in beta. Accepting a suggestion with Tab does not send it automatically, and generating predictions does not count toward usage limits during the beta.
Anthropic has opened Claude Code Projects to every Pro and Max subscriber on its waitlist. A four-minute walkthrough posted alongside the news advises keeping the coordinator at low reasoning effort to save usage.
Asana ran a 144-run study with GPT-6 Astra in Codex and brought the per-run cost of the StackAI browser agent down to $0.47. Most of the savings came not from switching models but from caching the browsing history and removing screenshots in batches.
Anthropic has opened dynamic workflows in Claude Managed Agents as a public beta. A program written by the lead agent runs up to 1,000 agents in phases within a single run and combines their results.
Sakana AI has released a paper presenting a benchmark that measures how well AI review systems catch contradictions inside research papers, along with MLR, a review-assistance system. MLR caught 73.43% of contradictions that overturn a paper's core claim, while existing systems stopped at 11 to 15%.
OpenAI has published a case study on British security company Sophos and its use of Daybreak. Average response time for cases handled by agents fell from about 38 minutes to 89 seconds, and AI now resolves 52% of MDR cases end to end.
OpenAI uncovered two Russian and Iranian false-front operations that hid behind a fake think tank and fake journalists, and banned the accounts involved. Russia's "Dark Clark" became the first operation to reach Category 5 on the breakout scale since OpenAI began reporting.
OpenAI published a case study on how LegalOn Technologies cut its Codex costs. LegalOn matched GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to task difficulty and split budgets by business, reducing estimated daily costs by about 65%.
