AI developer tools · What shipped

Gemini 4 Argon, Google's latest frontier model, is now globally available via API and multiple platforms. The model ranks #1 on the Vals Index benchmark across critical economic sectors including finance, coding, legal, and tax work. Key performance achievements include a 2.85x end-to-end inference speedup on Gemma 4 E2B via vLLM (reaching 108,000 tokens per second), 40% reduction in Python codebase session prep time, 2.3x speedup on quantum chemistry package xTB, and 87% score on the Embodied Reasoning benchmark. The model excels at complex tasks across coding, finance, legal, cybersecurity, and science domains, with demonstrated applications including Chrome architecture refactoring completed in weeks, Google's Agent Development Kit performance optimization, and real-world cybersecurity vulnerability discovery. Argon offers a 1M token context window via the Gemini API at introductory pricing of $2 per million input tokens and $10 per million output tokens.

Claude Haiku 5.5 was benchmarked on Artificial Analysis Intelligence Index v4.3.2, which incorporates 10 evaluations including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The model was evaluated across capability indexes for finance & accounting, strategy & ops, legal, healthcare & medical, engineering, and economics, with performance metrics covering intelligence scores, cost per task, and output speed in tokens per second.

Liquid AI has released Liquid Foundation Models (LFMs) designed for efficient inference across diverse hardware including GPUs, CPUs, and NPUs, from edge devices like wearables and phones to cloud infrastructure. The models emphasize speed, memory efficiency, and millisecond latency, with purpose-built decision models for classification and routing tasks. All LFMs are available free to download, run, and fine-tune commercially until companies reach $10M in annual revenue.

AWS published kernel engineering optimizations for Mixture-of-Experts (MoE) models on Amazon SageMaker AI, focusing on improving LLM inference performance. Three key optimizations were implemented: Swap-AB for WGMMA utilization, which transposes matrices to better align with Hopper GPU architecture and achieved 2-4 percent throughput improvement (upstreamed to vLLM 0.24.0+); backend routing for FP8 tensor quantization, which directs FP8-quantized models to the Triton backend instead of CUTLASS for 10 percent throughput gains (available in AWS LMI v28 container); and Triton JIT recompilation elimination, which removes unnecessary kernel recompilations by adding do_not_specialize annotations to variable-length parameters, yielding 9 percent faster time-to-first-token (TTFT) on multi-GPU setups (upstreamed to vLLM 0.22.0+). Combined, these optimizations deliver up to 10 percent higher throughput and 9 percent faster TTFT for MoE model inference.