AI News: Big Companies Push Agentic Coding, Small Devs Deliver Open Models
May 23, 2026 AI News Roundup — Big tech doubles down on autonomous coding agents while independent devs and open-source projects deliver powerful new models and tools. Here’s the most important developments from the past 24 hours.
Big Company Moves
Google Gemini 3.5 Flash Goes GA Google released Gemini 3.5 Flash as generally available. Optimized for complex agentic workflows and coding, it delivers 4x faster output tokens and strong benchmarks (76.2% Terminal-Bench 2.1, 83.6% MCP Atlas). Powers Search, Gemini app, and enterprise tools. Gemini 3.5 Pro expected next. Source
OpenAI Codex Advancements Continued rollout and improvements to Codex, the autonomous coding agent that handles full features, bug fixes, tests, and PRs in sandboxed environments. Early enterprise adoption strong. Related
Small Devs & Open-Source Highlights
Moonshot AI Kimi K2.5 Open Weights Moonshot released Kimi K2.5 as open weights — native multimodal model strong for agent workflows. Major win for accessible high-performance models.
Indie Coding Agent Tools New and updated open-source coding agents and frameworks gaining traction for local-first, repo-level work. Focus on practical automation for solo devs and small teams.
Key Papers & Research
- Continued work on agent evaluation, security, and long-horizon reasoning.
- AlphaEvolve-style evolutionary coding agents for scientific discovery remain influential.
Open-source momentum is accelerating access to frontier capabilities.
Sources & Further Reading
Organized for easy digestion. More updates tomorrow.