Annonces

OpenCode reached $40M ARR and 4.6M weekly users after an Anthropic Claude ban drove developers to its open-source AI coding agent, CEO Jay V said on a YC podcas

Anthropic's Claude Opus 5 nearly matches its flagship Fable 5 on benchmarks at half the price but deliberately weakens cybersecurity.

Microsoft's MAI models live in Bing and Office cut GPU costs 84-89% in production, altering AI cost math.

Alibaba's 0.8B OvisOCR2 end-to-end model beats every pipeline document parser with 96.58 on OmniDocBench.

Google's Gemini 3.5 Flash Cyber, a lightweight fine-tuned model locked to governments, found 55 V8 vulnerabilities vs Claude Opus 4.6's 36.

Hugging Face revamps its Inference Endpoints service to cut deployment complexity, offering vLLM, SGLang, and custom containers with autoscaling and pay-as-you-

OpenRTAG, a new benchmark from researchers, reveals how text-attributed graph models fail on dirty data, unifying nine degradation scenarios across nine dataset

PRO-LONG, a July 2026 framework, uses programmatic memory to let LLM agents recall 47 steps back, boosting ARC-AGI-3 pass rates by 18 points.

A new arXiv paper repositions LangGraph as a workflow orchestration tool, not an AI benchmark, with three executable recipes for stateful processes.

Voxtral TTS, a new hybrid text-to-speech model, beats ElevenLabs Flash v2.5 with a 68.4% win rate in voice cloning and releases its weights openly.

New tutorial from researchers reveals production failures of LLM agents not captured by academic benchmarks, with mitigations including verification pipelines a

PoTRE, a new AI framework using four specialized reasoning agents, achieved 49.92% on Humanity's Last Exam by proving diversity of thought outperforms sheer mod

Alibaba ATH's HappyOyster 1.0 builds a drivable 3D world from a single photo or text prompt, now in gray testing.

Researchers artificially lesioned an AI model to reproduce stroke-induced aphasia errors, matching individual patient profiles with high accuracy.

Qwen Cloud launched Token Plan Individual and Team tiers, bundling AI models to cut costs 40% and integrate with Claude Code and Cline.

Researchers built a neural network that learns emotional associations via Pavlovian conditioning, reproducing human-like preferences for images.

Google DeepMind's Gemma 4 crossed 300 million downloads, signaling open-weight models now outperform proprietary APIs on cost.

July 22 arXiv study shows natural-language autoencoders pass reconstruction tests despite 98% false claims; RECAP achieves AUC 0.95 for verifiable explanations.

OpenCode adds inclusionAI's Ling 3.0 Flash for free, betting token efficiency wins the AI coding war.

Sakana AI launched Fugu Ultra 1.1, a routing layer that hides which AI model processes your prompt via a single API to cut costs and complexity.