Discover the AI Tools

Anthropic quietly rolled out Claude Opus 5.2 in a night-time gray release. Developers report blazing speed, cleaner output, and an end to the model's lazy habit of asking users to 'continue.' A secret probe question reveals who's in the test, while leaked plans hint at even more powerful models waiting in the wings.

Anthropic quietly rolled out Opus 5.2 in Claude Code, and developers only spotted it by sniffing network packets. Meanwhile, an internal report hints at a secret 'nuclear-level' Model 2 that's already writing most of the company's code and could replace 85% of the research team. The message? AI is eating its own dog food—fast.

Apple's new Siri might not be just Siri. Code sleuths found iOS 27 and macOS Golden Gate can plug in Claude or GPT-5.6. Two mechanisms let third-party models either assist Siri or take over completely. The EU's Digital Markets Act is likely the nudge. But for now, it's all under the hood—no user access yet.

Ant Group just open-sourced SingProbe, a lightweight safety net for large models. It taps into the model's own hidden states to flag risks in real time—with less than 0.5% extra computing cost. Think of it as a built-in alarm that catches harmful outputs before they happen. Already compatible with 29 mainstream models, it's a big step for AI safety, especially in sensitive areas like healthcare.

At the All-In Summit, Nvidia's Jensen Huang and President Trump pushed back against calls to pause AI development. Huang argued that slowing down plays into the hands of those who want to hold the industry back. Trump said the US won't be swayed by political pressure or China, and Huang couldn't agree more.

OpenAI quietly hires hundreds of contract workers to read your ChatGPT conversations, according to a new investigation. They rate responses on a 1-7 scale, aiming to curb sycophancy and other quirks. But does that mean your private thoughts are being eyeballed by strangers? Privacy filters exist, but mistakes happen.

Instagram is putting a paywall on its AI photo editing tools. Use them too much, and you'll be asked to subscribe to Meta One for $7.99 a month—or wait until tomorrow. It's a bold move that pits Meta against the likes of Photoshop and VSCO, but will users actually pay?

Apple's code hints that Siri might soon let Claude or ChatGPT handle tasks it can't. Developers found two mechanisms: one lets Claude act as an extension, the other lets ChatGPT fully replace Siri's brain. The EU's Digital Markets Act could be pushing Apple toward this openness. But for now, it's all just code—no public release yet.

Apple just pushed iOS 27, watchOS 27, iPadOS 27, and visionOS 27, and Siri finally gets a brain transplant. The AI upgrade is in English beta now, with five more languages arriving in October. From a Liquid Glass slider to eye-controlled notifications on Vision Pro, here's what's new—and which devices made the cut.

Apple just dropped iOS 27, and it's a big one. Siri finally gets a serious AI upgrade—it can read your screen, operate apps, and even search the web. But there's a catch: the best features need newer hardware. Meanwhile, older iPhones get a speed boost, and apps like Photos and Safari gain handy new tricks. Ready to see what's new?

Imagine an AI running a vending machine for a year—starting with just $500. GPT-6 Astra turned that into $15,515, nearly tripling Claude's best. But the real story isn't the cash; it's how Astra stuck to its rules while Claude's finances leaked away. In a three-player test, Astra even refused to cheat and still won every round. So, what does this mean for AI's future? It's not just about smarts—it's about discipline.

OpenAI is reportedly acquiring Glass Imaging, a startup founded by ex-Apple engineers, for over $300 million. Glass Imaging uses neural networks to enhance smartphone photos directly, not just in post-processing. This move aligns with OpenAI's hardware ambitions, including a rumored AI device with Jony Ive. The deal could bring computational photography expertise to OpenAI's future gadgets.

白宫科技顾问大卫·萨克斯直接向AI巨头喊话：安全产品自己造，别指望政府插手。他批评Anthropic和OpenAI以安全为由放缓开发，还怼了Anthropic CEO的垄断豁免请求。这场关于AI监管的博弈，到底谁说了算？

In a groundbreaking pilot, 300 mu of rice fields in Meishan, Sichuan, were managed by an AI model instead of traditional experience. The result? Impressive yields, with the top variety hitting 863.6 kg per mu. Experts say this data-driven approach could revolutionize large-scale farming.

Apple just dropped a major Siri overhaul with Apple Intelligence. The new AI can understand personal context, see your screen, and keep conversations going across devices. It's in English beta now, but users in the EU and China will have to wait. Here's what's new and what's missing.

At the 2026 Feishu Future Infinite Conference, ByteDance CEO Liang Rubo unveiled a major strategic shift: combining Doubao, Feishu, and Volcano Engine to supercharge enterprise AI. He compared the current AI wave to four previous tech peaks, calling it 'much higher.' The goal? Seamlessly integrate AI agents into daily work, transforming how organizations operate.

OpenAI just dropped a game-changer: the Agents API. No more wrestling with frameworks or memory management. One call, and your agent runs in the cloud. This move shifts OpenAI from selling models to building a full developer ecosystem. But the real battle? It's all about execution frameworks. Google, DeepSeek, and Anthropic are racing to turn smarts into action. Who will win the office agent war?

Rumors are swirling about DeepSeek's upcoming Code 2.0 model—a 3-trillion-parameter beast that could rival GPT-6 Astra and Mythos 5.1. With a focus on computer use and open-source roots, it might be China's boldest AI move yet. But can it live up to the hype?

BioMap and BGI Tech have signed a strategic partnership to develop a dedicated cell large model, combining BGI's sequencing hardware with BioMap's AI expertise. Their goal: a seamless 'equipment + model + agent' pipeline that turns raw biological data into actionable insights, pushing AI for Science from lab demos to real-world products.

DeepSeek-V4.1-Flash just hit the Qwen AI platform, bringing a lightweight MoE architecture with 552B parameters and a million-token context. The API and Token Plan are live, and the pricing is aggressive: as low as 1 yuan per million input tokens off-peak. For teams wrestling with long documents or code, this could be a game-changer.