AI developer tools · What shipped

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, positioning them as the world's most advanced models for coding and knowledge work. Fable 5.1 achieves approximately 25% lower costs than Fable 5 for typical workloads through 75% reduced cache read pricing, with savings up to 45% for highly agentic tasks. The model demonstrates significant performance improvements across multiple benchmarks: Terminal-Bench 4.0 (55.8% vs 42.0% for predecessor), CursorBench 3.2.0 (73.4% vs 70.5%), and Humanity's Last Exam (65.0% with tools vs 63.8%). Fable 5.1 includes improved safeguards with 60% fewer false positives in cybersecurity while now supporting defensive vulnerability discovery, and enhanced biology safeguards that fire 85% less often for benign requests. Mythos 5.1 (identical underlying model with different safeguards) demonstrated advanced scientific capabilities including designing high-affinity protein binders at 10x higher binding affinities than competition entries, creating a new Venus elevation map with 2-3km detail, and optimizing deep learning models by up to 2.5x through custom GPU kernels. Both models are available immediately on Claude API and major cloud platforms.

OpenBMB has released MiniCPM5-2B, a 2.52 billion parameter dense language model with a 131,072 token context window and Apache 2.0 licensing. The model averages 53.9 across 34 benchmarks, outperforming Qwen3.5-4B at 51.1 and excels particularly on tool use tasks (97.1 on τ²-Bench Telecom), coding agents (69.1 on LiveCodeBench v6), and long-context retrieval (68.1 on NoLiMa). Training combines 400B tokens of supervised fine-tuning, reinforcement learning with specialized task teachers using the JustRL II algorithm, and on-policy distillation merging 16 RL experts into a single model. The model uses standard LlamaForCausalLM architecture and runs on mainstream inference engines including vLLM, SGLang, Transformers, llama.cpp, and Ollama without custom kernels. OpenBMB has also released all training datasets and intermediate checkpoints open-source, including UltraData-Code, UltraData-Math, specialized agent datasets with 500K samples, and RL training data with 80K+ samples.

DeepSeek V4 Flash Vision Exp has been released as open weights. Users are reporting benchmark results from the model running on DGX Spark hardware, with performance metrics ranging from 87-91 on a hardmode 88-question benchmark depending on configuration. The model supports up to 650k context windows and is being tested with multi-agent orchestration setups, though some users report occasional looping issues at high context limits that can be mitigated with adjusted thinking settings.

Syncfusion has released Agent Skills and MCP Servers as new AI coding assistant tools designed to improve the reliability of AI-generated code for Syncfusion components. Agent Skills are project-level guidance packages that provide component-specific implementation patterns, including setup, required modules, properties, events, theming, and accessibility guidance, allowing AI assistants to generate code following recommended practices. MCP Servers give AI coding assistants direct access to current Syncfusion documentation, APIs, and code examples through a #search_docs tool, enabling real-time lookup of recently released features and API changes without leaving the editor. Both tools are installable separately and work across multiple frameworks and development environments including VS Code, Cursor, and JetBrains IDEs.