Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
SemiAnalysis published the first official long-context, multi-turn agentic inference benchmark results on Google's TPUv7 (Ironwood) across four models on its InferenceX/AgentX benchmark, finding the chip either competiti…
A benchmark by Arun Prabhakaran of LatentAILabs measured a 10-turn DevOps chat on Qwen/Qwen3-8B in bfloat16 on an NVIDIA A100-SXM4-40GB and found that re-prefilling the full transcript processes 5,549 tokens to read a 96…
Anthropic's modified S-1 filing states that "we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI,'" a disclosure flagged in Zvi Mowshowitz's AI news…
A developer built DeadWave 13, a Halloween-themed interactive AI experience that pairs a "Ghost Radio" feature, where tuning fictional frequencies between 88.0 and 108.0 MHz maps to distinct ghost personas for LLM-genera…
A developer built Agent-Token-Optimizer, an open-source Python library for deterministic prompt compression and JSON pruning that reports 60% to 85% token reduction in under 0.5 milliseconds with no external dependencies…
Omnara's open-source agent harness matched the pass rates of Claude Code and Codex on Terminal-Bench 4.0 while running 11% cheaper and 12% faster, the company reported. With GPT 6.1 Sol, the Omnara harness cost 11% less …
OpenAI launched its next-generation image generation model, Images 2.6, on Wednesday, adding text-to-GIF creation, reusable reference images, and image creation up to 50% faster than before. Developers can choose between…
A developer released llm-evals, a Ruby CLI gem (version 0.5.0) for building multi-dataset, multi-case evaluation suites that compare models, effort levels and temperatures and rank them by a single weighted score balanci…
Liquid AI released LFM3, a family of open-weight models led by the 1.7B-parameter LFM3-1.7B, along with multimodal LFM3-VL-2B and LFM3-Omni-2B variants, built on a 24-layer tri-hybrid architecture that interleaves gated …
Amazon Marketplace built an A+ Content quality-evaluation pipeline on Amazon SageMaker AI that renders a full product detail page as a single snapshot for a vision-language model to grade holistically, then fuses the VLM…
Anthropic's AI models have been improving 1.9 times faster since October 2025 than previously, with AI itself accounting for perhaps 80% of the acceleration, according to an 11-week assessment by AI safety nonprofit METR…
OpenAI launched GPT-6 Cyber on Thursday, a version of GPT-6 Astra trained to find software flaws, build working exploits and write fixes, available only to vetted Daybreak Red customers through the API and Codex with no …
Baidu released ERNIE-5.5-Flash-Base, an open foundation model with 280 billion total parameters and 14 billion active per input, on Hugging Face in FP8 and BF16, deliberately shipped without fine-tuning or post-training.…
OpenAI's GPT-6 Cyber model scored 48.5 percent on its ExploitGym training evaluation, up from 32.1 percent for its predecessor, the company said. OpenAI said the model can detect and address zero-day vulnerabilities and …
A developer built AutoMode, a free MIT-licensed Claude Code plugin that routes each prompt to Haiku, Sonnet or Opus with a matching reasoning effort, running on the user's own Claude login without an API key or proxy. In…
A developer used the Reporails CLI's xray command to inspect a ten-month-old LangGraph project's Claude Code session and found that only two of 11 loadable instruction files reached the agent's context. The nested app/gr…
Google released Gemini Omni 1.2 Flash on November 5th, 2026, making the video generation and editing model generally available through the Gemini API and Gemini for business, with a stated 720p price of $0.07 per second.…
OpenAI announced GPT-6 Cyber, a security-research model that the company says pairs stronger capabilities with better model alignment, and said it is available to select Daybreak Red customers with access expanding over …
A Python-based source-audit harness that sends one brief and one numbered source pack to ChatGPT (GPT-6 family) and Claude (Claude 5.5 family), runs five trials per assistant, and uses deterministic code rather than a se…
Check Point Research found 2,557 systems hosting or having hosted ARTEX panels, with 1,608 live and publicly available instances at the time of its research, as threat actors used the Chinese open-source AI-powered penet…