AI developer tools · What shipped

Google released Gemini 3.7 Flash, a multimodal model improving substantially over Gemini 3.6 Flash with gains from 49% to 65% on DeepSWE and from 17% to 30.4% on Automation Bench, achieving an Intelligence Index score of 56 comparable to GPT-5.6 Terra at $0.75/$3.75 per million input/output tokens, half the previous pricing. Five frontier-level models entered the top 20 this week: SpaceXAI released Grok 4.6 scoring 61 on the Intelligence Index at $2/$6 per million tokens with 69.9% on CursorBench 3.2 and 61.3% on FrontierCode 1.1; Z.ai released GLM-5.3 outcompeting Kimi K3; Alibaba published Qwen 3.8 Max as open weights with 2.4T total parameters and 95B active parameters supporting 262,144-token native context; DeepSeek released V4 Pro 0813 with 1.7T total parameters and 49B active parameters scoring 87.9% on Terminal-Bench 2.1 and 62.7% on DeepSWE. Local AI model releases included Qwen 3.8-27B achieving 42% on DeepSWE 1.1 and 73% on Terminal Bench 2.1, Meta's Muse Glimmer 30B under Apache 2.0, Nvidia's Nemotron 3.5 Lightning 30B MoE model supporting one-million-token contexts, and Cohere's North Micro Vision Instruct 2.4B vision-language model. For developer tools, DeepSeek released DeepSeek Harness for agent execution accumulating 23,000 GitHub stars within days, SpaceXAI released Grok Bot for persistent AI agents available on desktop and iOS, Microsoft released MAI-Code-1.1-Flash integrated into GitHub Copilot with 25% greater token efficiency at one-quarter the cost, and Writer released Palmyra X6 reducing operating costs and improving speed by over 40%. Video and image generation updates included Lightricks' LTX-2.5 22B diffusion model for text-to-video and image-to-video supporting multi-shot sequences, Alibaba's Wan-Animate-2 14B for character animation, and MiniMax Music 3 for music generation.

Claude Fable 5 leads the OpenHands Index coding-agent benchmark with an 81.0% average score across five categories: issue resolution, frontend work, greenfield development, testing, and information gathering. The July 31, 2026 snapshot evaluated 33 AI model variants, with Claude Opus 4.8 (71.9%) and Claude Opus 4.7 Adaptive (69.7%) following in second and third place. OpenHands Index is currently displayed on BenchLM as a reference benchmark but excluded from overall model rankings, with quarterly refresh cadence to track performance changes across real-world software engineering agent tasks.

Composer 2 leads the React Native Evals benchmark at 96.1%, followed by Composer 2 Fast (94.9%) and GPT-5.4 (85.3%). The open benchmark evaluates AI coding agents on real-world React Native implementation tasks, with 16 models assessed across framework-specific mobile app development work including navigation, animation, and async state management. The benchmark was refreshed August 15, 2026 and operates on a quarterly refresh cadence, currently displayed as reference material with 20% weight in BenchLM's overall scoring system.