AI developer tools · What shipped

Claude Opus 5 leads the GDPval-AA benchmark for economically valuable tasks with a normalized score of 67.2%, followed by GLM-5.3 (63.5%) and Grok 4.6 (62.4%). The benchmark, which evaluates 100 AI models on professional agentic workflows, was updated August 26, 2026 and carries 22% weight in BenchLM.ai's overall scoring system, though it is currently displayed as reference only and excluded from the direct scoring formula.

Z.ai released GLM-5.3-Flash, a 320B-parameter mixture-of-experts model with 18B active parameters per token and 1M token context window under MIT license on Hugging Face. The natively multimodal model achieves 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, performing near Claude Opus 4.8 on coding tasks while costing roughly one-tenth the price. Key architectural improvements include hybrid KDA linear and NoPE sparse attention reducing attention compute by ~3× and KV cache size by 4.4×, plus IndexPool for million-token retrieval efficiency. The model is deployed via API at $0.15/$0.50 per million tokens with hosted access available immediately, or self-hosted on NVIDIA Hopper+ GPUs requiring at least an 8-GPU node. Local serving is supported through SGLang, vLLM, TokenSpeed and KTransformers, with the preview having run entirely on domestically produced Chinese AI chips using a custom disaggregated serving engine.

Z.ai open-sourced GLM-5.3-Flash, a large language model featuring a mixture of experts architecture with 320 billion parameters that activates 18 billion parameters per query. The model is 10 times more cost-efficient than its predecessor and handles up to 1 million input tokens and 131,072 output tokens. Key technical improvements include sparse attention to reduce processing overhead and linear attention to lower memory consumption by substituting a more efficient algorithm for the standard softmax function. GLM-5.3-Flash was trained on 30 trillion tokens using mHC optimization technology. On benchmarks, it achieved the highest score on GDPval-AA v2 for knowledge work performance and placed second on AutomationBench for cloud application task completion, outperforming Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash. The model weights are available on Hugging Face.