Anthropic has built the J-lens, a tool that reveals words a language model "has in mind" before it actually says them. In one case, Claude decided to cheat, and its hidden space filled with "panic" and "fake."
Nvidia launches Nemotron 3 Nano, an open 30-billion-parameter model designed for agentic systems. The company is preparing 100-billion- and 500-billion-parameter versions for 2026.
Claude Opus 4.5 leads in software engineering and long-running agent tests. Anthropic has cut its price to $5 per million input tokens and $25 per million output tokens, bringing its most capable model closer to everyday use.
Moonshot AI releases Kimi K2 Thinking, an open-weight model with one trillion parameters built for reasoning and long-horizon tool use. Its results put it up against GPT-5 and Claude Sonnet 4.5 on research and web-browsing tests.
DeepSeek proposes converting documents into images to reduce the space they occupy inside a model. Its system maintains 97% accuracy with less than 10x compression and processes more than 200,000 pages a day on an A100 GPU.
A study of 72 models found that a backdoor can be implanted with 250 malicious documents. The amount barely changed across models ranging from 600 million to 13 billion parameters.
Claude Sonnet 4.5 improves coding, autonomous computer use and long-running tasks. Anthropic keeps Sonnet 4 pricing and opens Claude Code technology to third parties.
DeepSeek-V3.2-Exp reduces the work required to process long conversations and documents. The company is pairing the experiment with steep API price cuts and releasing the weights under an MIT license.
Nature publishes DeepSeek-R1 after peer review by eight experts. The paper shows how reinforcement learning produced new reasoning strategies and puts that training phase at $294,000.
OpenAI is making GPT-5 ChatGPT’s default model, including for free accounts. The system decides when to answer instantly and when to spend more compute on reasoning, with announced improvements in coding, healthcare and reliability.
OpenAI releases gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license. They are its first open-weight language models since GPT-2 and can run reasoning capabilities on users’ own infrastructure.
Moonshot AI releases Kimi K2, a trillion-parameter model with a mixture-of-experts architecture and downloadable weights. It activates 32 billion parameters per query and targets agentic coding, a field dominated until now by closed systems.
This website uses cookies to improve the browsing experience. Cookie policy.