In September 2026, over 40% of pull requests created on GitHub started with a first draft written by an AI agent, up from 17% in March. This figure appears in the annual edition of the Octoverse report, which also confirms TypeScript’s leading position among the platform’s languages.
Record activity on GitHub from June onwards
Pull requests are proposed code changes submitted for project review before being merged. The report measures GitHub’s activity, not the entire development market. In the twelve months ending September 30, 2026, its key metrics at least doubled year-over-year, combining public and private repositories: 17.9 billion pushes versus 6.7 billion in 2025, 34.3 billion commits versus 12.9 billion, 45.4 billion GitHub Actions jobs versus 14 billion, and 1.3 billion merged pull requests versus 538 million.
The growth accelerated sharply from June onwards, to the point that the monthly commit volume in September 2026 reached six times that of September 2025. GitHub links this inflection point to the arrival of more capable models and agents able to follow a task over several days, without establishing a causal link between the two. The cited 2025 values have been recalculated according to the new edition’s rules: the platform now includes private repositories and runs its year from October to September. They therefore differ from those published last year, and only comparisons internal to the 2026 edition are valid.
Agents at the heart of pull requests and model choice
In September, 110 million pull requests were initiated by an AI agent, six times more than in March. GitHub specifies this is a floor, as only agents it can detect are counted. These requests are also broader: 87% touch multiple project files, compared to 79% for others. Their acceptance rate has improved: in February, 83% were merged into the code within two weeks, versus about 86% for others, while in September, they reached 90%. ‘For a growing share of the work, the first move is now to direct an agent,’ writes GitHub.
Among GitHub Copilot users who choose their own model, 61% used multiple models in September, up from 50% four months earlier. Paying subscribers used 47 distinct models that month, versus 29 in May. In the week ending October 13, 38% of requests went to generalist models, like Claude Sonnet 5 or GPT 6.1 Sol, 37% to reasoning and planning models, like Claude Opus 5.5 or GPT 5.5, and 25% to fast and lightweight models, like Claude Haiku 5.5 or Gemini 3.8 Flash. No family clearly dominates. Many subscribers also let Copilot choose for them: 63% activated automatic selection at least once in September.
TypeScript leads, projections to 2027, and ever more developers
The language podium remains unchanged between 2025 and 2026. In monthly contributors, TypeScript gained 3.62 million contributors, Python 3.23 million, and JavaScript 1.82 million. In repositories created in 2026, TypeScript is the main language in 35% of those where an agent was detected, versus 24% of others. The report describes an association, not an observed choice by agents. GitHub’s model average, published without a range, places TypeScript at 13.07 million monthly contributors in September 2027 and Python at 12.56 million.
GitHub claims over 225 million developers on its platform. The United States (38.3 million) and India (32 million) remain in the lead, and France has 4.75 million, ranking 10th globally. For the first time, the report also ranks public repositories by sector, based on an estimate established by an AI model on samples: construction accounts for about 9% of relevant repositories, the highest share of all sectors. Of the 43 forecasting models GitHub ran up to 2031, 21 anticipate at least a tripling of the number of developers.
