What AI models can do, and what has meaningfully changed. Articles are ordered by original publication date.
Google says Gboard uses a new way to train next-word suggestions on protected servers. Its privacy claims have conditions and remain unverified by this review.
DeepSeek's experimental V4-Flash-Vision-Exp adds native image input to the low-cost Flash API tier at unchanged per-token pricing for multimodal agent workflows.
An NYU Stern preprint finds the strongest model tested produced reusable algorithms competitive with the best specialized methods on three OR benchmark families.
12-model study: pooled commitment on unknowable questions rose from 6.5% to 54.0% as relevant-looking evidence was added — fabricated panels matched genuine ones.
Cohere Parse (parse-v5.0) turns enterprise documents into structured Markdown at $1.50 per 1,000 pages, making document parsing an infrastructure cost decision.
Qwen explicitly frames Qwen3.8-Flash-Next as a preview of Qwen4 architecture: open-weight multimodal MoE with 125B params, 6B active per token, QSA, and N-gram embeddings.
A Nature Biotechnology benchmark found AI-designed antibodies statistically indistinguishable from experimentally optimized results — with important caveats.
Z.ai's GLM-5.3-Flash brings MIT-licensed open weights and Z.ai-reported gains of 17–22 pp on coding and agent benchmarks at roughly one-ninth of GLM-5.2's API price.
ServiceNow researchers report 20–35-point gains across three enterprise agent benchmarks by evolving the harness while keeping model weights fixed.
AI models can find the right fact in a 128K-token document and still not use it. New research shows why retrieval alone doesn't validate an AI analyst.
NVIDIA's AVO reaches 100% on ARC-AGI-3's public set with Claude Opus 5. Why the vendor-reported result matters for evaluating agent systems.