Aug 13, 2026

xAI's new model (Aug 12), built for long-running agents and more ambitious interactive and visual work. API pricing is $2 in / $6 out per 1M tokens (a faster variant at double the cost). Claims 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol.

Why it matters: The debate centers on catching the leaders on measured intelligence while undercutting them on price. A separate thread formed around how to read the benchmarks (strongest on agentic tasks).

Alibaba releases open weights for a 2.4T-total / 95B-active MoE on Hugging Face, claiming 92.6 on GPQA Diamond and 67.7 on SWE-bench Pro, with local deployment supported via vLLM and SGLang.

Why it matters: At 2.4T total parameters this is an outlier in scale for open weights, and the thread centers on how far the gap to closed frontier models has narrowed — plus early notes on what hardware it actually takes to run.

Modular's systems language for AI reaches 1.0, with a stability pledge that 1.x changes will be primarily additive. One language targets CPUs, GPUs, and accelerators, and Modular commits to open-sourcing the compiler and toolchain in 2026.

Why it matters: The attempt to rewrite the AI infrastructure layer outside Python-plus-CUDA just crossed the stable-release line. With substantial outside contribution already in the standard library, the thread debated whether it sticks as a language.