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 adding nine external benchmarks covering agentic work, cybersecurity, algorithm engineering, forecasting, and research-level physics.

Tencent officially launched Hy3, a hybrid fast-and-slow-thinking model with 295 billion total parameters and 21 billion active parameters supporting 256K context length. The model demonstrates performance comparable to flagship models with two to five times larger parameter scales while achieving greater stability and cost efficiency than its preview version. Hy3 has been integrated into Tencent products including WorkBuddy/CodeBuddy, Yuanbao, Marvis, and ima, with particular improvements in code generation, complex reasoning, and agent capabilities for productivity tasks like software development and financial modeling. The model is available under Apache 2.0 license on Hugging Face and ModelScope, with progressive rollout to third-party developer platforms including OpenRouter, Hermes, Kilo, Cline, and others. Hy3's API is now accessible on Tencent Cloud TokenHub, and token consumption has increased twentyfold since the preview launch, reflecting market adoption for practical cost-effective applications.

DeepSeek V4 Flash, a 284B parameter Mixture-of-Experts model with 13B activated parameters supporting 1M-token context, leads OpenRouter's coding model rankings with 5.49T tokens used, optimized for fast inference and high-throughput workloads with strong reasoning and coding performance. Xiaomi's MiMo-V2.5 ranks second with 4.55T tokens, delivering Pro-level agentic performance at roughly half the inference cost with native omnimodal capabilities, while MiniMax-M3 uses sparse attention replacing full attention with KV-block selection to reduce per-token compute to roughly 1/20 the previous generation cost at 1M tokens. Z.ai's GLM 5.2 reasoning model and other recent releases from Tencent, Anthropic, and StepFun also rank highly, with models featuring 1M-token context windows, configurable reasoning levels, and optimization for agentic workflows and complex multi-step software engineering tasks based on July 2026 real usage data from developers.

AWS has released a production-ready infrastructure for deploying multi-turn reinforcement learning with Amazon Nova Forge on Amazon SageMaker HyperPod. The solution provides an event-driven pipeline that automatically provisions compute, routes rewards, and runs training when data is uploaded to S3, featuring a two-phase deployment model that separates long-lived foundational resources from ephemeral per-run resources. The infrastructure uses SageMaker HyperPod (EKS) for training with GRPO weight updates, ECS on Fargate for reward workers, and the Nova Forge SDK for message routing between the model and environment, with full orchestration via AWS Step Functions and EventBridge. A sample repository with AWS CDK deployment code is available for adapting the Wordle placeholder environment to custom RL tasks and agent workflows.