AI-powered programming agents have crossed a threshold few would have predicted a year ago: they are no longer auxiliary tools, but the primary development engine in the companies that adopt them.
Three data points from this week confirm it from different angles.
80% of Code at Anthropic Is Already Written by Claude
On June 4, 2026, the Anthropic Institute published “When AI Builds Itself,” an internal analysis revealing that over 80% of code merged into production at Anthropic was written by Claude. Before the launch of Claude Code in February 2025, that figure was in the single digits.
The numbers are hard to ignore: Anthropic engineers ship 8 times more code per day than in 2024. Claude’s success rate on open-ended engineering tasks jumped from 26% to 76%. In one documented case, Claude applied more than 800 fixes in April 2026 that reduced a class of API errors by a factor of 1,000 — work an engineer estimated would have taken four years to do manually.
There is an important caveat: Anthropic is an AI lab whose main product is AI tools. Its workflow is optimized for this. This metric is not directly extrapolable to the entire industry. Anthropic itself acknowledges that code written by Claude “was somewhat worse than human-written code” in certain dimensions, and that the engineer’s role has shifted from writing code to directing and reviewing.
But the direction is clear.
OpenCode: 7.5 Million Developers and 160,000 Stars
While Anthropic published its internal metrics, the open-source project OpenCode reached 160,000 stars on GitHub and 7.5 million monthly active developers. Its official site (opencode.ai) promotes two key features that explain its adoption: support for over 75 LLM model providers through Models.dev, and LSP integration that automatically loads the correct language servers for the code being edited.
OpenCode works with local models (Ollama, llama.cpp, vLLM) and cloud APIs (Claude, GPT, Gemini), making it a flexible, free alternative to Claude Code and Cursor. Its model-agnostic architecture is likely the main reason for its explosive growth.
MiniMax M3: Sparse Attention and 1M Context
The third data point comes from the model side. MiniMax launched M3, an open-weight model with three innovations that put it on the map for coding agents: 1 million tokens of context — with a guaranteed minimum of 512K — a proprietary sparse attention architecture (MSA) achieving 9.7x acceleration in prefill and 15.6x in decode at 1M tokens, and native multimodality.
The HuggingFace community describes it as “the first open-weight model to combine top-tier programming performance, a million-token context window, and native multimodality.” On programming benchmarks, it competes directly with Claude Opus 4.7 and GPT-5 Codex, but at a fraction of the cost.
The Pattern
All three points converge on the same thing: coding agents are no longer a curiosity. They are infrastructure. Claude Code demonstrates real productivity at enterprise scale. OpenCode democratizes access with an open, multi-model platform. MiniMax M3 pushes the technical limits of what an open model can do.
For a developer, the question is no longer “should I use a coding agent,” but “which one serves me best for what I need.”
Sources: Anthropic Institute — When AI Builds Itself · OpenCode · MiniMax M3 · HuggingFace — MiniMax Goes Sparse