Artificial Analysis Finds 1 in 4 AI Endpoints Quietly Underperform Their Advertised Models
A revamped benchmark reveals that nearly 1 in 4 serverless inference endpoints quietly lose accuracy versus the official model weights.
20 stories
A revamped benchmark reveals that nearly 1 in 4 serverless inference endpoints quietly lose accuracy versus the official model weights.
Antigravity moves agent execution into Google-managed cloud VMs, letting coding tasks run in the background across GitHub repos and Drive files from any device.
A Stanford team wired DSPy into HELM and found that swapping one static prompt for structured ones lifts scores 6% and reshuffles rankings.
Moonshot AI's 2.8-trillion-parameter open-weight model lands on Azure's model catalog with 1M-token context, vision, and aggressive agentic benchmark scores.
Thinking Machines releases Inkling 2 Small, a 594B-parameter MoE with 17B active that scores 31 on the AA Intelligence Index.
Black Forest Labs' 5.6B open-weights model lands within one Elo point of Qwen-Image-2.1 on text-to-image while costing a fraction of larger rivals.
A Vietnamese sentence embedding model fusing SimCSE contrastive learning with PhoBERT has quietly crossed 850K downloads on Hugging Face.
Cursor now turns a prompt into a deployed web app with a Postgres database, auth, and security scans behind an SSO link.
Krea bundles Sora 2, Veo 3.1, Seedance 2.5 and Kling 3.0 into one browser workspace with camera controls, presets and sound generation.
Ampersand's multi-agent 'Co-Coding' mode hits 92.2% on SWE-bench Pro HARD-51 without any coding-specific training, matching frontier models at 70% lower cost.
World Labs just put Atlas, their autoregressive world model, into a browser demo that streams generated frames from cloud GPUs at up to 15 fps.
An independent audit of Anthropic's internal R&D pipeline finds AI model progress has roughly doubled since late 2025, mostly driven by AI itself.
Meta's new audio-driven Diffusion Transformer streams 25fps avatar video at sub-second latency, beating Runway and HeyGen in head-to-head ratings.
Microsoft Research's PDD Kit uses Verus formal methods to mathematically prove agent-generated Rust code matches developer intent, replacing exhaustive line-by-line review.
A new framework lets quadrupeds walk safely on rough terrain using only IMUs, encoders, and contact sensors, cutting position error 64.8%.
Anthropic adds a read-only money hub to Claude that connects to US bank accounts through Plaid and answers spending questions in plain English.
Runway plugs Google's newest video model into its editor, offering 15-second clips, conversational edits, 4K upscaling, and sharper shot-to-shot consistency.
Nanbeige4.3-33B is a dense 33B open-weights reasoning model that matches larger peers on agentic benchmarks and runs locally under Apache 2.0.
Decart's Lucy 2.6 sharpens character, clothing, and product fidelity in live-streamed AI video edits, shipping at the same price as 2.5.
Goodfire's new framework uses a model's internal representations as training signals to kill reward hacking, sycophancy, and spurious shortcuts at the source.