AI developer tools · What shipped

Claude Fable 5 achieved a new high score of 161 on the Epoch Capabilities Index, surpassing GPT-5.5 Pro by 1 point and marking the first time Anthropic has led the index in over a year. Epoch AI recently expanded its benchmarking hub by tracking 13 new evaluations, with 7 incorporated into the Capabilities Index, while also adding nine external benchmarks covering agentic work, cybersecurity, algorithm engineering, forecasting, and research-level physics as of late June 2026.

Transformers library v5.13.0 released with major model additions including Kimi 2.5-2.7 (multimodal agentic model with advanced coding capabilities across Rust, Go, Python), MiMo-V2-Flash (27T token MoE model supporting 256K context), Nemotron 3.5 ASR (600M multilingual speech model with configurable latency tradeoffs via 80-1120ms chunks), Qwen3 ASR (Whisper-style encoder with Qwen3 decoder supporting automatic language detection), ZAYA1 (760M active / 8.4B total MoE model with compressed convolutional attention), VideoPrism (video encoder achieving state-of-the-art on 31/33 benchmarks across question answering and vision tasks), RADIO (NVIDIA vision foundation models via multi-teacher distillation), and MiniCPM3 (4B dense model outperforming 7B-9B models using Multi-head Latent Attention). Breaking changes include standardized layer declarations for ONNX/torch.export/ExecuTorch exportability and fixed attention masking for Gemma 3/4 models.

EvoPolicyGym, introduced in an arXiv paper on July 2, presents a new benchmark for evaluating autonomous coding agents through iterative policy evolution. The framework separates visible training feedback from hidden validation and held-out test cases, enabling agents to learn from failed attempts in a controlled environment with server-mediated rollouts and strict budgets. GPT-5.5 achieved top performance across the Core-16 benchmark suite, and the GitHub repository provides infrastructure, protocol documentation, and adapters for multiple agent implementations including Claude Code and Codex CLI, offering a pattern for long-running agent evaluation that emphasizes exploration-exploitation tradeoffs and reproducible feedback loops.