Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Google's Gemini embedding models underwent a major pricing restructuring between 2024 and 2026, with the cost per million tokens increasing tenfold from $0.02 for the legacy text-embedding-004 to $0.20 for the new gemini…
An engineer building a benchmark for AI coding agents discovered that an LLM judge incorrectly scored a half-finished audit as 'exhaustive' because it lacked a reference answer. The judge evaluated the plain agent's audi…
OpenAI released GPT-4o on May 13, 2024, expanding ChatGPT adoption. The model has been benchmarked on math, multimodal, and reasoning tasks, with scores including 48 on math, 62.2 on multimodal grounded, and 2.7 on Human…
Jua, a Zurich-based physics AI company, uses ClickHouse Cloud to power its physics foundation model EPT-2, cutting forecast delivery time from one hour to 20 minutes and reducing compute costs by a third. The platform de…
GitHits and Context7 are competing tools that provide AI coding agents with external context about libraries and dependencies, with GitHits offering additional features like source code navigation and implementation exam…
Doubleword released an open-source agent swarm that lets an LLM orchestrator design its own team of bounded-context workers to parallelize wide, shardable tasks. In a security audit of a 2.4M-token codebase, the swarm ou…
Google's Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) is now available on Vercel's AI Gateway, offering fast, low-cost image generation at $0.034 per 1K images. The model generates images alongside text in under 4 se…
SigNoz, Langfuse, and Arize Phoenix lead the top LLM observability tools in 2026, offering capabilities such as full-stack monitoring, agent workflow debugging, and drift detection. The tools converge on OpenTelemetry's …
Hugging Face and the EvalEval Coalition launched an integration that allows contributors to submit standardized evaluation results (EEE schema) to Hugging Face Community Evals, consolidating scattered benchmark scores in…
Anthropic's Claude Code now supports autonomous scheduled workflows by combining a custom /goal command with routines and cron scheduling, allowing developers to set finish conditions and run recurring tasks like nightly…
Apple, Anthropic, and OpenAI are racing to connect AI assistants to users' real-world context—files, calendars, codebases—because raw model intelligence is no longer the main bottleneck for utility. The context gap, wher…
DeepSpark's confidence-scheduled verifier reduces GPU waste in AI agent inference by skipping verification for low-probability tokens, dynamically adjusting thresholds under load to trade minimal accuracy for significant…
OpenAI Codex introduces a record-and-replay feature that lets users automate repetitive computer tasks by demonstrating a workflow once, with AI reasoning making it more flexible than traditional RPA tools. The system in…
Stanford's STORM method uses five expert agent personas to produce research 25% more organized than single-prompt approaches. The multi-agent workflow assigns different perspectives to interrogate a topic, resulting in m…
Anthropic's Claude Code supports two multi-agent architectures: sub-agents, which operate in isolated sessions under a central orchestrator and cannot communicate with each other, and agent teams, which enable peer-to-pe…
DeepSeek released DeepSpark, an open-source speculative decoding system that accelerates LLM inference by 50–400% without retraining. The method uses a small draft model to propose multiple tokens in parallel, which a la…
ZhipuAI released GLM 5.2, an open-weight large language model with a 1-million-token context window and MIT license, which outperforms GPT-4.5 on design and coding benchmarks. The model offers frontier-level performance …
Anthropic's Claude Sonnet 5 is now available on Vercel AI Gateway, offering improved coding and agentic performance at Sonnet pricing. The model features enhanced instruction following, document parsing, and long-context…
Researchers from an undisclosed institution introduced AlignAtt4LLM, the first application of the AlignAtt policy to decoder-only large language models for simultaneous speech translation at IWSLT 2026. The system, a syn…
Researchers evaluated four state-of-the-art zero-shot cross-lingual voice cloning systems across six languages, finding that Arabic remains particularly challenging. The study, presented at IWSLT 2026, analyzed tradeoffs…