AI developer tools · What shipped

JetBrains released the Kotlin Benchmark, an official benchmark for evaluating AI coding agents on Kotlin software engineering tasks based on SWE-bench methodology. The benchmark includes 105 engineering tasks sourced from active open-source repositories, with solutions verified in containerized environments. Initial results show Claude Code with Opus 4.7 achieved the highest resolution rate at 85.71% (90 of 105 tasks), followed by JetBrains Junie with Opus 4.7 at 81.9% and Codex with GPT 5.5 at 81.9%. The benchmark assets and methodology are publicly available on GitHub, with a leaderboard tracking evaluation results at kotlinlang.org/benchmark. Future iterations plan to expand ecosystem coverage, add evaluation metrics beyond test passing, and evaluate additional agents and model configurations.

Claude Fable 5 achieved a new high score of 161 on the Epoch Capabilities Index, surpassing GPT-5.5 Pro by 1 point and marking the first time Anthropic has led the index in over a year. Epoch AI recently expanded its benchmarking hub by tracking 13 new evaluations, with 7 incorporated into the Capabilities Index, while also adding nine external benchmarks covering agentic work, cybersecurity, algorithm engineering, forecasting, and research-level physics.

Claude Fable 5 debuted as Anthropic's first generally available Mythos-class model in July 2026, entering the WebDev AI Leaderboard at 1653 Elo—92 points clear of second place, the widest gap ever recorded. It features 1M context window, 128K output, always-on adaptive thinking, and costs $10/$50, making it the most expensive model in the comparison. Claude Opus 4.8 simultaneously entered at #2 with 1561 Elo at unchanged $5/$25 pricing, offering improved long-horizon agentic performance and 2.5x output speed in fast mode. On the tools side, OpenCode held the #1 position with 160K+ GitHub stars and 7.5M monthly active developers, maintaining model-agnostic access to 75+ providers with unique LSP integration feeding compiler diagnostics back to the model. Claude Code moved meaningfully stronger by gaining access to both Fable 5 (highest Elo in any tool) and Opus 4.8, plus computer use capabilities enabling direct terminal-based browser navigation and app interaction. SpaceX announced a $60 billion acquisition of Cursor expected to close Q3 2026, giving Cursor access to xAI's Colossus infrastructure for model training but raising trust concerns as Cursor's market share dropped from 41% to 26% while Anthropic's climbed toward 50%.

OpenRouter's coding model rankings updated July 2026 show real usage data from developers. Top performers include DeepSeek V4 Pro (1.6T parameters, 49B activated) excelling at advanced reasoning and software engineering benchmarks with 1M-token context, Claude Opus 4.8 delivering strong multi-step reasoning and complex coding performance, and Tencent's Hy3 (295B MoE, 21B active) optimized for production agentic workflows with configurable reasoning modes. Other notable releases include DeepSeek V4 Flash for cost-efficient coding assistance, MiniMax-M3 with sparse attention reducing long-context compute by roughly 1/20 compared to prior generation, and Step 3.7 Flash with 196B parameters and selectable reasoning levels for coding and structured outputs. Models support 256K to 1M token context windows across specialized applications from code generation and debugging to full-codebase analysis and multi-stage project orchestration.