Anthropic Says Claude Now Writes Over 80% of Its Shipped Code

·1 views

Anthropic revealed in a June 2026 essay that more than 80% of the code it ships is now written by its AI model, Claude, up from low single digits just two years ago. The shift was accelerated by Claude Code, a tool that allows the model to autonomously read codebases, make edits, run tests, and fix errors. Human engineers have moved from writing code to reviewing and approving the model's output, with each engineer reportedly shipping roughly eight times more code per quarter than before. Beyond volume, Anthropic says an unreleased internal model now outperforms its own researchers at choosing research directions, and nearly closed the gap with human experts on an unsolved AI safety problem. However, all key figures come from Anthropic's own internal, unreleased models, meaning the claims have not yet been independently verified.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

ZCode: A Claude Code Alternative Launched by GLM Creators

ZCode is a new AI coding tool released by the team behind GLM, positioning itself as an alternative to Anthropic's Claude Code. The product was announced and is accessible via the Z.ai platform. It has drawn early attention on Hacker News, garnering initial points and community discussion. Details about its full feature set and pricing remain limited at this stage.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Self-Speculative Decoding Cuts AI Reward Training Time Without Quality Loss

Researchers have introduced a technique called self-speculative decoding to speed up the reward-based fine-tuning phase of AI model training, where models repeatedly generate answers to practice and improve. The method creates a compressed, lower-precision copy of the model at each training step to quickly draft text, while the full model only verifies those drafts rather than generating every word itself. Because the clone is rebuilt from the live model at every step, it stays in sync with the constantly changing training model and avoids accuracy drift. The system also intelligently disables speculation when hardware is already at full capacity, activating it only when spare resources are available. The final trained model is identical in quality to one trained without the technique, making the speedup effectively lossless — a notable claim in a field where efficiency gains are often overstated.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer finds dead-code bug in own AI security scanner while probing LLM vulnerabilities

A developer built AgentProbe, a tool that fires 49 known attack prompts across 8 categories at AI models to test their resistance to prompt injection, currently ranked the top security risk for LLM applications by OWASP. While building the scanner, the developer discovered a logic bug where a custom 'hedge-then-comply' detector always returned a confidence score of 1, but the escalation threshold was set at 2 or higher, meaning the detector's results were silently discarded every time. As a result, every case the cheap keyword detector was meant to handle was unnecessarily escalated to a more expensive LLM-as-judge call, wasting resources and creating a single point of failure. The bug went unnoticed because the LLM judge independently caught the same patterns, masking the fact that the keyword stage was effectively dead code as a decision-maker. The incident highlights a broader concern in AI evaluation: LLM-as-judge systems are widely used in safety benchmarks and model leaderboards, yet the reliability of the judge model itself is rarely verified.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DomainShuttle AI Model Tackles Subject Consistency in Text-to-Video Generation

Researchers introduced DomainShuttle, a new AI video generation model, on June 27, 2026, targeting a long-standing challenge in text-to-video synthesis. The model aims to keep a specific character or object visually consistent across frames while still allowing natural, dynamic motion. It achieves this through a panel of specialized 'temporal experts,' each handling a different aspect of motion, which are dynamically combined based on the scene's needs. An improved spatial-temporal tracking mechanism further helps maintain subject coherence through complex movements. The approach has attracted early community interest via a public code repository, with potential applications in personalized content, advertising, and entertainment.

0 comments Read more at DEV Community