AI developer tools · What shipped

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

A German research consortium has released Soofi S 30B-A3B, an open language model that achieves top benchmark scores among fully open models in both English and German, surpassing OLMo 3 32B and Apertus 70B. The 30-billion-parameter mixture-of-experts model activates only 3.2 billion parameters per token, using a hybrid Mamba-Transformer architecture that maintains throughput at very long contexts—generating roughly eight times more tokens per second than dense 14-24B models at 40,000-token context lengths. Trained on 27 trillion tokens with deliberately weighted German data (7.2% in phase one, 15.3% in phase two), Soofi S scores 73.8% on HumanEval and 70.2% on MBPP code benchmarks, leading all fully open models on German benchmarks by double-digit margins in most categories. The training ran March-May 2026 on up to 512 Nvidia B200 GPUs at Deutsche Telekom's Industrial AI Cloud in Munich, and the consortium is releasing model weights, checkpoints, complete training code, and detailed data inventory documentation.

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released "Inkling," a 975 billion parameter open weights model under Apache 2.0 license. The model is the largest American open weights model to date, comparable in size to Chinese models like DeepSeek V4 and GLM 5.2. It features a mixture of experts architecture with 256 routed experts and 2 shared experts, trained on 45 trillion tokens using Nvidia GB300 NVL72 systems, and supports a million-token context window. Inkling matches Nvidia's Nemotron 3 Ultra on Terminal Bench 2.1 using roughly one-third the thinking tokens, with efficient chain-of-thought reasoning capabilities. The model is available via Thinking Machines' Tinker platform for customization and fine-tuning, downloadable from Hugging Face, and compatible with multiple inference engines including vLLM, SGLang, and Llama.cpp. A companion model, Inkling-Small at 276 billion parameters with 12 billion active parameters, is also in preview. The company is integrating the model with third-party API services including TogetherAI, Fireworks, Modal, Databricks, and Baseten.

Google's Gemma 4 family of open models includes six variants: 12B Unified, E2B, E4B, 26B-A4B, and 31B. These multimodal models support 140+ languages, up to 256K context, offer both dense and MoE variants, and are Apache-2.0 licensed. The 12B Unified features unified text, image and audio support and runs on 8GB RAM (4-bit) or 14GB (8-bit), while the 26B-A4B MoE model balances speed and accuracy with 4B active parameters, and the 31B offers strongest performance for maximum quality. Unsloth has released optimized GGUF and MLX quantizations with dynamic 4-bit and 8-bit options, published KL Divergence benchmarks showing their quantizations on the Pareto frontier, and offers inference via llama.cpp, Ollama, and their open-source Unsloth Studio web UI. Performance benchmarks show the 26B-A4B achieving 89.2% on AIME 2026, 80.0% on LiveCodeBench v6, and 2150 Codeforces ELO, with multimodal capabilities including 76.9% on MMMU Pro and vision benchmarking across OmniDocBench and MATH-Vision.