AI developer tools · What shipped

v2.1.239 includes cost estimation updates for data-residency workspaces, Alpine/musl build support for native tools like image paste and clipboard, improved Python migration support (claude-api upgrade for anthropic 0.x to 1.x migration), and multiple reliability fixes including Bedrock streaming fixes, proxy handling improvements, and session management corrections. v2.1.238 adds keybinding customization (readline flavor support), plugin marketplace headers helpers for authentication, and self-hosted runner improvements. v2.1.237 fixed prompt caching for LLM gateways and custom base URLs, plus added a "Concise" output style. v2.1.236 introduces ANTHROPIC_DEFAULT_MODEL environment variable, cross-session idle notifications, and macOS sandbox precedence rules for file access. Recent versions also address infrastructure stability including memory management in long sessions, cloud session resilience, and remote control connection reliability improvements.

Google has released Gemini 3.7 Flash, a new AI model optimized for developer workflows and coding tasks. The model achieves 65.3% on DeepSWE v1.1 (up from 49.0% for 3.6 Flash) and 43.6% on FrontierCode 1.1 Main (up from 34.4%), demonstrating significant improvements in code generation accuracy and multi-step agentic task execution. Key features include better instruction following, tool calling, autonomous error recovery, and web development capabilities with an Elo score of 1588 on LMArena's WebDev Arena. The model introduces flexible thinking levels (low, medium, high) to balance latency and intelligence, and is now the default across Google Antigravity, Google AI Studio, and Gemini Enterprise Agent Platform, priced at $0.75/1M input tokens and $3.75/1M output tokens introductory pricing through December 31, 2026.

Martian, a frontier AI research team, partnered with Daytona to scale agent training infrastructure for ARES, an open-source reinforcement learning framework for LLM coding agents. By integrating Daytona's sandbox platform, Martian achieved 95% faster execution of SWE-Bench Verified benchmarks and provisioned 20,000+ parallel sandboxes in one week, compressing evaluation times from hours to minutes and saving three months of engineering effort. The integration enables agents to run real coding tasks with real-time feedback at scale, with each sandbox dynamically provisioned to match task dependencies and immediately torn down after use to maintain efficiency.