AI developer tools · What shipped

As of July 12, 2026, the MATH-500 benchmark leaderboard shows GPT-5 leading at 99.4%, followed by o3 and Grok 3 Mini at 99.2%, with 111 models evaluated overall. The benchmark measures performance on competition mathematics problems requiring multi-step reasoning across algebra, geometry, number theory, and calculus. Notable recent releases include Claude Sonnet 4 Thinking at 99.1%, Grok 4 at 99.0%, and o4 Mini at 98.9%, while specialized thinking models from Alibaba Qwen3 and others continue advancing mathematical reasoning capabilities alongside pricing comparisons across providers.

Moonshot AI released several AI coding and machine learning infrastructure projects relevant to developer tools and benchmarks. Kimi-Dev-72B, an open-source coding LLM, achieves 60.4% performance on SWE-bench Verified for issue resolution. Kimi Code is a fast AI coding agent for development workflows. On the infrastructure side, Moonshot released Mooncake, a KV-centric disaggregated LLM serving framework that won Best Paper at FAST 2025, and checkpoint-engine, middleware for efficiently updating model weights in LLM serving. Additional releases include Kimi K2 (32B activated parameters with state-of-the-art coding performance), MoBA for long-context LLMs, and Moonlight optimizer for improved LLM training efficiency.

OpenAI released the GPT-5.6 model family with three versions: Sol (flagship), Terra (balanced, everyday use), and Luna (lower-cost). Sol achieved state-of-the-art scores on coding benchmarks including 72.7% on DeepSWE, 92.2% on BrowseComp, and 62.6% on OSWorld 2.0, with significantly improved token efficiency compared to prior models. OpenAI also launched GPT-Live-1 and GPT-Live-1 mini, full-duplex voice models enabling simultaneous listening and speaking, and introduced ChatGPT Work, an agentic interface integrating developer tools with unified plugins and faster computer use capabilities. SpaceXAI and Cursor launched Grok 4.5, a 1.5 trillion parameter mixture-of-experts model trained on real agent-interaction data, achieving Opus 4.8-level coding scores: 64.7% on SWE Bench Pro and 62% on DeepSWE. Meta released Muse Spark 1.1, a multimodal reasoning model with frontier-level performance on coding and computer use tasks, featuring a one-million-token context window and costing $1.25/$4.25 per million tokens. Additionally, Cognition released SWE-1.7, a coding model achieving 42.3% on FrontierCode benchmarks, and OpenAI published a report finding 30% of SWE-Bench Pro tasks are broken, suggesting the benchmark ceiling is near 70% and recommending reliance on alternatives like DeepSWE.

Google Cloud released Cloud Run sandboxes in public preview, enabling secure, isolated execution environments for AI-generated code within serverless instances. This allows developers to safely run dynamically generated scripts from LLMs without leaving the Cloud Run environment. Additionally, Google Cloud introduced the Google Cloud Workbench Notebooks extension for VS Code, allowing data scientists and developers to run ML notebooks on managed cloud environments directly within their local IDE, streamlining the ML development lifecycle. The company also announced Gemini 3.1 Flash-Lite, a cost-efficient model optimized for high-volume developer workloads, and Gemini 3.1 Pro, a more capable baseline for complex problem-solving, both available in preview via Vertex AI and the Gemini API. Cloud Run worker pools reached general availability for pull-based, non-HTTP workloads with queue-aware autoscaling via the open-sourced Cloud Run External Metrics Autoscaler (CREMA). Furthermore, Google Cloud's new AI Telemetry Collector agent standardizes TPU monitoring using OpenTelemetry, and Run:ai Model Streamer now supports TPUs with Google Cloud Storage in TPU vLLM 0.18.0, achieving over 2x faster model loading while halving peak memory usage.