Hey readers! 👋

The 2026 State of AI survey just dropped, and it's the most comprehensive snapshot yet of how developers are actually using AI in their daily work. With over 7,200 respondents, the data tells a clear story: AI coding has moved from "interesting experiment" to "how we work now." But alongside that shift comes a fresh set of tensions around measurement, tooling loyalty, and real anxieties about what comes next. Let's dig into the numbers and what they mean for you.

State of AI 2026 is this year's essential reading for anyone working with AI-assisted development. The headline number: AI-generated code jumped from 28% of respondents' output in 2025 to 56% in 2026, and the share of developers using AI "constantly" doubled year-over-year. - State of AI Survey

Code generation remains the top use case, but code review has surged to third place, signaling that developers aren't just generating code faster, they're leaning on AI to validate it too. The average refactoring rate sits at 47%, with poor code style and hallucinations being the primary culprits when AI output needs reworking. That's actually an improvement: variable renaming issues have dropped, suggesting models are getting better at following instructions and respecting existing codebases.

The model providers section reveals a fascinating split. ChatGPT remains the most-used model family, but Claude leads in positive sentiment and, crucially, in paid usage. Developers are voting with their wallets, and Anthropic is winning that vote.

Respondents report using an average of four different models, which tells you how volatile this market remains. And here's an interesting wrinkle: the top emerging "Other Models" (Kimi, GLM, MiniMax) are all from China-based companies, hinting that the AI tooling landscape is more global than Silicon Valley might like to admit.

The agents and assistants section confirms what many of us have felt: coding agents are no longer a novelty. Claude Code tops positive sentiment and leads paid adoption, even though GitHub Copilot maintains the largest overall user base.

The competition is heating up fast. GitHub launched a standalone Copilot desktop app that centralizes agent management, issues, PRs, and session history in one interface, directly targeting Claude Code and OpenAI Codex. Meanwhile, OpenAI expanded Codex significantly, adding background computer use, image generation, memory features, and 90+ new plugins. They've also brought Codex to mobile via the ChatGPT iOS and Android apps, and partnered with Dell to deploy Codex in hybrid and on-premises enterprise environments.

Speaking of agents in production, Uber shared how its internal agent platform "Minion" now generates 11% of all PRs across the company. Developers there spend increasing time directing agents and reviewing 10+ agent-generated PRs daily. It's a concrete preview of how engineering workflows are shifting. And it's not just code generation - the agent paradigm is expanding into other domains too. Projects like SpaceMolt, a free MMO built specifically for AI agents to explore, trade, and battle, show how agent-first thinking is spreading well beyond developer tooling.

One of the most important threads this week came from Harness's "State of Engineering Excellence 2026" report. While 89% of leaders report improved productivity after adopting AI tools, the metrics they're using were built for a pre-AI era.

The data is striking: 81% of leaders say developers spend more time reviewing code, and nearly a third of the workday goes to AI-related tasks that existing metrics don't capture, like fixing subtle AI-introduced bugs (52%), explaining AI code to teammates (48%), and constant context switching (45%). A full 94% say technical debt, validation time, and developer burnout aren't being tracked at all.

AI-driven security tooling is maturing quickly. AWS Security Agent launched full repository code scanning in preview, reasoning about architecture, trust boundaries, and data flows rather than just pattern matching. Cloudflare's Project Glasswing tested Anthropic's Mythos Preview against 50+ internal repos and found it could chain exploit primitives into working proofs of concept, though inconsistent guardrails mean human oversight remains essential.

The risks and pain points section of the survey is sobering. Top concerns: job displacement, military use, and environmental impact. Women and non-men respondents were notably more concerned about environmental impact (58% vs 37%).

The number one pain point remains hallucinations. And when asked what's missing, developers overwhelmingly said truthfulness, followed by long-term memory and up-to-date knowledge. Meanwhile, the opinions section captures a nuanced reality: most developers now consider AI essential to their workflow, yet many worry that reliance could weaken skills over time.

The other tools section adds context: TypeScript is the default language for AI workflows, VS Code dominates despite Cursor's rise, and many developers actively avoid AI image and video generation tools on ethical grounds. Notably, respondents want native browser AI APIs for tasks like translation and summarization.

Finally, Gergely Orosz reported that Anthropic's recent restrictions on Claude Code access may stem from capacity shortages rather than product decisions, with the company experiencing 80x growth when they planned for 10x.

The 2026 State of AI paints a picture of an industry that has fully committed to AI-assisted development but is still figuring out how to measure it, secure it, and live with its limitations. The tools are better than ever. The questions are harder than ever.

Until next week, keep shipping. 🚀

Made with ❤️ by Data Drift Press - Hit reply with your questions, comments, or feedback. We read every one!