Aider blog

Separating code reasoning and editing

An Architect model describes how to solve the coding problem, and an Editor model translates that into file edits. This Architect/Editor approach produces SOTA benchmark results.

SEP 26, 2024

o1-preview is SOTA on the aider leaderboard

Preliminary benchmark results for the new OpenAI o1 models.

SEP 12, 2024

Sonnet seems as good as ever

Sonnet's score on the aider code editing benchmark has been stable since it launched.

AUG 26, 2024

LLMs are bad at returning code in JSON

LLMs write worse code if you ask them to return the code wrapped in JSON via a tool function call.

AUG 14, 2024

Coding with Llama 3.1, new DeepSeek Coder & Mistral Large

Summary of code editing skill for the new models, with Sonnet and GPT-3.5 for scale.

JUL 25, 2024

Sonnet is the opposite of lazy

Claude 3.5 Sonnet can easily write more good code than fits in one 4k token API response.

JUL 1, 2024

Aider is SOTA for both SWE Bench and SWE Bench Lite

Aider sets SOTA for the main SWE Bench, after recently setting SOTA for the Lite version.

JUN 2, 2024

Aider has written 7% of its own code

Aider has written 7% of its own code, via 600+ commits that inserted 4.8K and deleted 1.5K lines of code.

MAY 24, 2024

How aider scored SOTA 26.3% on SWE Bench Lite

Aider achieved this result mainly through its existing features that focus on static code analysis, reliable LLM code editing, and pragmatic UX for AI pair programming.

MAY 22, 2024

Linting code for LLMs with tree-sitter

Aider now lints code after every LLM edit and automatically fixes errors, using tree-sitter and AST-aware code context.

MAY 22, 2024

Drawing graphs with aider, GPT-4o and matplotlib

Use GPT-4o to draw graphs with matplotlib, including adjusting styles and making visual changes. You get the graph, but you also get the code in your repo.

MAY 13, 2024

Aider in your browser

Aider has an experimental browser UI, allowing you to collaborate with LLMs on code in your local git repo.

MAY 2, 2024

GPT-4 Turbo with Vision is a step backwards for coding

OpenAI's GPT-4 Turbo with Vision model scores worse on aider's code editing benchmarks than all the previous GPT-4 models. In particular, it seems much more prone to "lazy coding" than the existing GPT-4 Turbo "preview" models.

APR 9, 2024

Claude 3 beats GPT-4 on Aider's code editing benchmark

Claude 3 Opus outperforms all of OpenAI's models on Aider's code editing benchmark, making it the best available model for pair programming with AI.

MAR 8, 2024

The January GPT-4 Turbo is lazier than the last version

The new `gpt-4-0125-preview` model is quantiatively lazier at coding than previous GPT-4 versions, according to a new "laziness" benchmark.

JAN 25, 2024

Unified diffs make GPT-4 Turbo 3X less lazy

GPT-4 Turbo has a problem with lazy coding, which can be signiciantly improved by asking for code changes formatted as unified diffs.

DEC 21, 2023

Speed benchmarks of GPT-4 Turbo and gpt-3.5-turbo-1106

This report provides a detailed comparison of the speed of GPT-4 Turbo and gpt-3.5-turbo-1106 models based on the aider benchmarking suite.

NOV 6, 2023

Code editing benchmarks for OpenAI's "1106" models

A quantitative comparison of the code editing capabilities of the new GPT-3.5 and GPT-4 versions that were released in Nov 2023.

NOV 6, 2023

Building a better repository map with tree sitter

Tree-sitter allows aider to build a repo map that better summarizes large code bases.

OCT 22, 2023

GPT code editing benchmarks

Benchmarking GPT-3.5 and GPT-4 code editing skill using a new code editing benchmark suite based on the Exercism python exercises.

JUL 2, 2023

Improving GPT-4's codebase understanding with ctags

Using ctags to build a "repository map" to increase GPT-4's ability to understand a large code base.

MAY 25, 2023