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 also recently expanded its benchmarking hub by tracking 13 new evaluations, with 7 incorporated into the ECI, while adding nine external benchmarks spanning agentic work, cybersecurity, algorithm engineering, forecasting, and research-level physics.

Faros released an AI model routing evaluation framework that uses rubric-based scoring to assess AI coding model performance on real engineering tasks. The framework evaluates patches against weighted checklists of task-specific criteria rather than relying solely on test execution, enabling evaluation of work where original environments or complete test coverage may be unavailable. Validation against SWE-bench datasets showed gold patches scored a mean of 0.80, repeated judge runs had correlation of 0.96 with ±0.02 variation, and the framework achieved AUC rankings of 0.75-0.86 in distinguishing successful from failed patches. The framework demonstrated a Spearman correlation of +0.94 between model success rates and average rubric scores on their failed attempts, indicating it preserves meaningful quality distinctions that binary pass/fail benchmarks cannot capture.

Cohere released Tiny Aya, an open-weight multilingual model supporting 70+ languages designed for local deployment on resource-constrained devices. The Expedition Tiny Aya research program produced multiple projects with released benchmarks and datasets: Tiny Aya Math Edition created a public multilingual Math Olympiad benchmark for evaluating mathematical reasoning; Kids Companion developed a multilingual safety benchmark for child-focused AI systems across dozens of languages; Cross-Lingual Word-Sense Disambiguation introduced new resources for multilingual semantic evaluation; and Language-Aware Quantization explored compression techniques preserving performance across diverse languages. Additional releases include Tiny Aya Vision (parameter-efficient multimodal extension), Tiny Facade (4-bit quantized variants for on-device tool use), and open-source interpretability tools using sparse autoencoders for analyzing multilingual representations. The core Tiny Aya research was accepted to COLM 2026, with multiple teams publishing papers, benchmarks, datasets, and winning entries in the Hugging Face Build Small Hackathon.