All guidesUsing the tools

Four language charts, two winners, and how well AI codes in each

The popularity indices split between Python and JavaScript because they count different things. The year's moves are clearer, and the newest per-language AI test says that for mainstream languages the model is no longer the deciding factor.

6 min read17 September 2026

Every month someone posts a chart of the most popular programming languages, and every month the replies say it is wrong. The replies are half right. Four well known indices are in wide use, they count different things, and in September 2026 they split two apiece between Python and JavaScript. This piece sets the four side by side, picks out what actually moved in the last year, and then looks at the question we get asked more often than either of those. How well does AI code in each language, and why does it vary?

Four indices, four ways of counting

TIOBE counts hits for each language name across 23 sites including Google, Microsoft, Wikipedia and Amazon, and publishes monthly. September 2026 has Python first on 17.8% of hits, then C on 10.3% and C++ on 8.7%.
PYPL measures the share of Google searches for language tutorials, worldwide, and also publishes monthly. September 2026 has Python on 52.1%, Java on 13.3% and C and C++ together on 8.0%.
RedMonk ranks languages by GitHub pull requests and Stack Overflow questions, twice a year. The January 2026 edition, published in April, has JavaScript first, Python second and Java third.
GitHub's Innovation Graph publishes, for each of 184 economies and each quarter, how many developers pushed code in each language. We added the rows up for the first quarter of 2026. JavaScript comes out at 5.7 million, Python at 4.1 million and TypeScript at 2.8 million. A developer who pushes in two countries or two languages is counted in each, so this is our sum of GitHub's data and not a figure GitHub publishes itself.

Read together, the four agree more than they seem to. Python leads wherever people search and learn. JavaScript leads wherever people push code and ask questions. Both are true at once, and neither index is wrong.

What moved in twelve months

Rankings barely move. The year on year changes are more telling.

TypeScript. Developers pushing TypeScript on GitHub nearly doubled in a year, up 98% in our sum, the fastest growth of any major language. GitHub's own Octoverse report in October 2025 said TypeScript had overtaken Python and JavaScript by monthly contributors, and credited type systems for catching model mistakes before production.
Python. Down 8.2 percentage points on TIOBE as smaller languages picked up share, while Python pushers on GitHub rose 64% and it holds first place on TIOBE and PYPL. Falling share and rising use are not a contradiction. The pie grew.
Rust. Into the TIOBE top ten, at tenth, from eighteenth a year earlier.
C. Swapped places with C++ to take second on TIOBE. It is over fifty years old and still gaining share.
Go. Dropped from eighth to twelfth on TIOBE. It also scores lowest of the nine languages in the AI test below.

How well AI codes in each language

The standard test for AI coding is SWE-bench. A model is given a real issue from a real open source repository and has to produce a fix that passes the project's own tests. The best known version, SWE-bench Verified, is 500 Python issues. SWE-bench Multilingual, released in March 2025, applies the same method to nine other languages.

The newest per-language results any lab has published came with Anthropic's Claude Opus 4.5 announcement in November 2025. The chart carries no printed numbers, so we measured the bars against the axis and rounded to the nearest point. Java 91, C 84, Rust 81, JavaScript and TypeScript 80, PHP 73, C++ 72, Ruby 67, Go 63. For Python we used the same model's 80.9% on SWE-bench Verified, the same test family, as printed in its system card. The chart's error bars are roughly five points either way, so treat the ordering as the finding rather than the exact figures.

Three reasons the scores differ

Three things explain most of the gap between languages, and none of them is about the language being harder in any deep sense.

How much of the language the model has read. The Stack, the public code collection most coding models train on, holds about 64 GB of Python and about 1 GB of OCaml. A 2024 study published at OOPSLA found that models score roughly in proportion to that. The base model it tested solved 31% of Python problems and 7% of OCaml problems. Every language in the chart above has plenty of public code. Rare and in-house languages do not.
Where the model is tested. SWE-bench Verified is Python only, so Python was the language every lab tuned for first. Once a nine-language test existed in March 2025, scores on it rose from 43% to 78% in eleven months. What gets measured gets improved.
How fast the language says no. A compiler and a test suite tell the model it is wrong before a person has to, which is one reason Java, C and Rust sit at the top of the chart. It is not the whole story, because C++ and Go are compiled too and sit near the bottom. GitHub credits TypeScript's rise partly to its types catching model mistakes, and a 2025 study at PLDI showed that constraining a model to type-correct output cut TypeScript compile errors by more than half.

What we would do with this

For the mainstream languages the model is no longer the deciding factor. Every language on the chart is fixed correctly most of the time by a model that is now ten months old, and the gap between the top and the middle is smaller than the gap between any of them and a rare language. So choose for your team and the systems you already run, and let the tooling follow.

Python stays our default for business automation, because the libraries, the examples and the connectors are already there, and because it is the language every model has read most. For anything with a user interface we lean towards TypeScript, where the type checker does part of the review for you. If your estate runs on something rarer, AI can still help, but budget for more human review and expect more wrong answers. The model has read less of your language, and nothing in its training tells it when it is wrong.