AI developer tools · What shipped

Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, marking significant advances in AI model capabilities for coding and knowledge work. Fable 5.1 achieves approximately 25% lower costs for typical workloads through 75% reduction in cache read pricing (up to 45% savings for agentic tasks), while delivering superior performance across multiple benchmarks including Terminal-Bench 4.0, CursorBench 3.2.0, and Humanity's Last Exam. The model demonstrates substantial improvements in agentic coding, scientific research, and multidisciplinary reasoning, with Fable 5.1 scoring 55.8% on Terminal-Bench 4.0 versus Fable 5's 42.0%, and 73.4% on CursorBench 3.2.0. Mythos 5.1, the safety-configured variant, shows particularly strong performance in computational biology optimization, accelerating seven open-source deep learning models by up to 2.5× through custom GPU kernel optimization, reducing estimated GPU costs by 30-60% on genome-wide analyses. Both models are available immediately on Claude API and all major cloud platforms (AWS, GCP, Azure), with Mythos 5.1 restricted to US organizations through verified cybersecurity and life sciences access programs.

Google Research published "Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion" at ICML 2026, introducing the Retrieve-for-Train framework that optimizes AI search inference through offline reinforcement learning. Instead of expensive test-time reasoning, the framework trains a lightweight 53.9M-parameter diffusion model to generate diverse, complementary query results in a single non-autoregressive pass, achieving 12-20x speedup over autoregressive approaches while maintaining sub-second latency at production scale. The method uses composite reward optimization (groundedness, diversity, alignment) to train fan-out language models on open-source 4B models (Gemma3-4B, Qwen3-4B), then distills learned behaviors into the diffusion retriever, eliminating the need for chain-of-thought reasoning tokens and addressing inference bottlenecks in set-valued retrieval tasks across multimodal domains including fashion and music.

NVIDIA released a performance blog post detailing benchmarks for DGX Spark hardware running common AI tools, demonstrating how the system enables intensive AI tasks. The forum includes discussions of inference optimization techniques, including discussions around model quantization (Int4-AutoRound, NVFP4), inference engines (DGPP, a GB10-optimized C++/CUDA inference engine), and inference performance metrics (throughput ranging from 70-80 tokens per second with various model configurations like Qwen, DeepSeek, and GLM variants on single and dual DGX Spark setups).