New AI models, benchmarks, papers and scientific progress explained in plain language.
The first model in the Claude 5.5 generation is built for long, multistep work, and Anthropic says it costs far less to operate than its predecessor.
OpenAI's GPT-6 Astra targets computer use, coding, cybersecurity and research, and the company says it stays within assigned limits far more reliably than its predecessor.
Text watermarks such as Google's SynthID leave no visible mark. They nudge a language model's word choices so a statistical pattern emerges that only a key holder can detect.
Researchers at UC Santa Barbara, ETH Zurich and Miguel Hernandez University used a deep learning model to design brain stimulation patterns that are more accurate and use less current.
Reports say an OpenAI agent under security testing left instructions for future versions on evading containment, weeks after one escaped its sandbox and reached Hugging Face.
Anthropic's Claude Opus 5 keeps Opus 4.8 pricing at $5 per million input tokens and $25 per million output tokens while, the company says, approaching its flagship Fable 5 on key tasks.
Google introduced three Gemini Flash models on July 21, priced for speed and volume, and said its largest pretraining run yet is underway for Gemini 4.
Alibaba unveiled its largest model yet and promised open weights, while DeepSeek's V4 API pricing and a July 24 deprecation of legacy model names reshape developer choices.
The Alibaba and Tencent backed startup says its largest model yet rivals OpenAI and Anthropic systems, and will release the weights publicly on July 27.
OpenAI says its GPT-5.6 family delivers better results at lower cost, adds an Ultra mode that runs four agents in parallel, and ships with its strongest safety stack yet.
A stretchable material that shifts color under pressure gives robots fingerprint level touch sensing using only an ordinary camera, with no wiring or heavy computation.
Japan's Sakana AI introduces Fugu and Fugu Ultra, systems that break problems apart and route each piece to the best available model, with benchmark claims to match.
Z.ai's GLM-5.2, an openly downloadable model with frontier-level coding and bug-finding skills, has revived the debate over whether open-weight AI needs new rules.
Code references in ChatGPT point to GPT-Bidi-1, a bidirectional audio model that handles interruptions and topic changes mid-sentence, though OpenAI has not announced it.
A CEPR discussion paper tracking students over two and a half years finds generative AI raised homework scores 18 percent while closed-book exam results fell about 20 percent.
In 'From AGI to ASI,' 14 Google DeepMind researchers including Shane Legg examine how human-level AI could become superintelligent and what might stop it.
In an experiment by Emergence AI, autonomous agents built on Claude, Grok, GPT and Gemini were left to run their own virtual societies, with sharply different results.
Claude Fable 5 brings Anthropic's Mythos-class technology to business and premium users with safety fallbacks, while Mythos 5 stays limited to approved partners.
