Updated daily488 verified stories
Models News tracker.
New frontier, open-weight, multimodal, reasoning, and specialist AI models—plus the evidence behind their capability claims.
- 38stories this week
- 488sources cited
- 488verified timeline entries
Verified sourcing
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
Plain English
What happened, why it matters, and what to watch — without the jargon.
No filler
When the signal is thin, we publish nothing rather than padding the feed.
More stories
9 storiesProduct
Anthropic publishes open reference blueprint for commerce AI agents
Anthropic has published a public Apache-2.0 repository with shopping and merchant agents built on Claude, including runnable retail, travel, telecom and entertainment examples.github.comInnovation
SAGE proposes a lower-cost way to evaluate task-oriented AI dialogue agents
A new arXiv preprint introduces SAGE, an evaluation system that checks whether each turn in a task-oriented AI conversation correctly advances the underlying workflow state. The authors report that its core version matched or exceeded evaluated LLM judges across four benchmark slices while using symbolic rules and…arxiv.orgInnovation
RestoreBench tests whether AI agents can restore power-flow convergence
A new arXiv benchmark evaluates chatbot, single-agent and multi-agent LLM systems on diagnosing and correcting non-convergent power-flow cases across two power grids.arxiv.orgInnovation
Gated-memory routing aims to reduce the cost of multi-agent LLM systems
A newly submitted paper proposes retaining only useful reasoning steps when multiple language-model agents collaborate. Its abstract reports higher average benchmark accuracy and a 31.9% reduction in HumanEval inference cost compared with the strongest baseline.arxiv.orgSecurity
Study finds AI deception signals transfer unevenly between test settings
An arXiv study of Llama-3.1-70B-Instruct reports that spontaneous and instructed deception share some internal direction, but detection and steering methods transfer asymmetrically between them.arxiv.orgInnovation
GLANCE speeds up vision-language model decoding by drafting token blocks in one pass
A new arXiv paper introduces GLANCE, a speculative-decoding method that uses a vision-language model’s fused state to draft whole token blocks at once. The authors report lossless greedy outputs and up to 2.93× faster decoding in tested workloads.arxiv.orgProduct
Claude Fable 5.1 is available in GitHub Copilot
GitHub announced Claude Fable 5.1 for eligible Copilot plans, with gradual rollout and data-retention conditions that administrators should review.GitHubProduct
Blockchain.News reports Lyria 3.5 launch in Google’s Gemini app
Blockchain.News reports that Google launched Lyria 3.5 in the Gemini app and API, with tracks of up to three minutes and integrations across several Google products. The report’s claims have not been independently confirmed here.blockchain.newsInnovation
BenchMIRT finds LLM benchmarks can mix safety and reasoning signals
Ai2 introduces BenchMIRT, a method that analyzes individual benchmark questions to identify which capabilities they measure. Testing 100 LLMs across 16 benchmarks, the researchers found that some evaluations combine safety and general-reasoning signals and that smaller question sets can often preserve much of the…huggingface.co
One useful briefing each week
Keep up with AI without living in the feed.
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Reach people who are learning AI
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Post an AI jobSubmit an AI tool