# Codex makes fewer bugs, but more people use Claude > Source: > Published: 2026-07-07 16:24:51+00:00 Report # State of AI coding 2026 AI models wrote 90% of the code that cubic agents reviewed last quarter. Which one produced the fewest bugs? We're in a good position to answer. Cubic reviews thousands of commits a day for teams like Resend, n8n, and Daytona, and many of them track AI attribution with our CLI. So for every commit, we know which model wrote it, and whether we found bugs in it. Benchmarks tell you what a model can solve; this data tells you what it breaks. The short version - GPT-5.5 writes the cleanest code we see. So far that hasn't changed what developers use: 80% of developers code with Claude Code, and most of the commits we review come from Opus models.[*](https://www.cubic.dev/inside-ai-coding#note-on-the-data) ### TL;DR GPT-5.5 produces fewer bugs per lines changed than any other model in our dataset. Claude Code remains the most popular coding agent, used by 80% of developers. The Opus family is the most-used set of models: 80% of developers run an Opus model as their primary model each week. Claude Fable was the fastest-adopted model in cubic's history: 30% of developers were using it within 48 hours of release. ### Model quality: GPT-5.5 produces fewer bugs than any other model Each new model produces fewer bugs than its predecessor (Opus 4.7 generated half the bugs of Opus 4.6), with one exception: Opus 4.8 shows a higher bug rate than 4.7. We don't have a firm explanation. One candidate: developers throw their hardest tasks at a new flagship in its first weeks, which inflates early bug counts. GPT-5.5 produces fewer bugs than flagship Anthropic models and is the least buggy model. ### Developers merge 3× more code than a year ago The typical developer has roughly doubled the number of merged PRs over the past year, from about 6 per month in mid-2025 to around 12 today. Those PRs are also 2.2× larger at the median. In summary, developers are merging roughly three times more code than they were a year ago, a substantial increase in developer output. AI tends to generate more verbose and boilerplate-heavy code than humans, but although the impact is difficult to measure precisely, the value delivered has clearly increased with the adoption of AI. Developers merged a median of 5820 lines of code changes in May 2026, compared to 1930 in June 2025, a year ago. That's a 3x increase. PRs merged per developer doubled in the past year, from a median of 6 per month (June 2025) to 12 (May 2026). PRs are getting bigger too, from a median of 56 lines of code changed per PR in June 2025 to 124 in May 2026, a 2.2x increase. ### Coding agents: Claude dominates Claude Code leads at 80% of developers. Cursor sits at 18% and loses share every week; Codex holds steady at 17%. The percentages sum past 100 because many developers run more than one agent. ### Model usage: new frontier models win Anthropic's Opus model family is clearly the most popular among developers, with 80% of developers using it as their primary model every week. OpenAI's GPT family comes in second and has seen increased adoption following the release of GPT-5.5. The week-over-week data reveals a clear pattern: new flagship frontier models are quickly adopted within two weeks of release. ### The short-lived story of Fable The purple blob in the model-share chart is Claude Fable. Anthropic released it on Tuesday, June 9. By Thursday, 30% of developers on the platform had used it, the fastest adoption of any model in cubic's history. Then it vanished. On Friday of that week, the U.S. government ordered its withdrawal, and Anthropic pulled the model. Even with only half a week of availability, it captured 15% of primary-model share. ### Are humans writing code at all? Among teams that track attribution with the cubic CLI, AI now authors the vast majority of code (90–100%), up from 78% in mid-April. These are the most AI-forward teams in the industry, so treat this as a leading indicator, not an industry average. You can track AI coding activity and attribution for your team using the [cubic CLI](https://docs.cubic.dev/ide/cli-review). * The dataset is based on hundreds of thousands of commits made by teams of at least 2 developers. These metrics focus exclusively on commits where AI attribution was tracked via our CLI, representing only a fraction of all commits. We started tracking this data around mid-February. While this doesn't represent the total codebase, it offers a compelling directional look at emerging patterns.