Anthropic Data: Over 80% of Code Written by Claude, AI Accelerates Self-Evolution

Internal data shows task durations doubling every 4 months. Over 80% of code is now authored by Claude. Recursive self-improvement — AI designing more powerful successors — may arrive faster than expected.

Anthropic released internal data revealing that AI is accelerating its own development at unprecedented speed. Quarterly code output per engineer has increased 8× compared to 2021–2025 levels. AI task horizons have expanded from a mere 4 minutes to 12 hours, doubling every four months. Over 80 percent of the company's code is now written by Claude, and benchmarks including SWE-bench and CORE-Bench have saturated within two years and fifteen months respectively. The trajectory points toward recursive self-improvement — AI systems autonomously designing more capable successors — which, while not yet realized, may be approaching faster than most researchers anticipated. As of May 2026, Claude authors the vast majority of Anthropic's software, fundamentally reshaping the development pipeline and raising urgent questions about the pace of AI advancement.

NVIDIA Launches Nemotron 3 Ultra: 550B MoE Open Model Built for Agents

A 550B total, 55B active hybrid Transformer-Mamba MoE with 1M context delivers 5× inference speedup and 30% lower cost for agentic tasks.

NVIDIA released Nemotron 3 Ultra, a frontier open model designed from the ground up for long-running autonomous agents. The hybrid Transformer-Mamba mixture-of-experts architecture packs 550 billion total parameters with only 55 billion active, achieving 5× faster inference and up to 30% lower cost on complex agentic workloads. With a 1-million-token context window and support for multi-token prediction speculative decoding, the model targets coding, research, and enterprise workflows requiring sustained reasoning, planning, and tool use over extended horizons. Weights are available under an open license on Hugging Face, with day-zero serving support from vLLM.

NVIDIA Unveils Cosmos 3: The First Universal World Foundation Model

Cosmos 3 understands and generates across text, image, video, sound, and action — a unified framework for physical AI.

NVIDIA introduced Cosmos 3, the world's first open omnimodel designed for physical AI. It ingests and produces across five modalities — text, images, video, audio, and action — providing a single unified framework. Built atop a new breakthrough architecture, Cosmos 3 aims to serve as the foundational layer for robots, autonomous vehicles, and any system that must first understand the physical world before acting within it. The model is open and available for research and commercial use.