Aider Review 2026: Still the Industry's Benchmark, But Who's Actually Maintaining It?
Correction, Oct 4, 2026: an earlier version counted 1,883 open issues plus ~500 pull requests; GitHub's 1,883 already included the pull requests. The split is about 1,400 issues and 500 pull requests. We also corrected the leaderboard size: it lists 69 models, not 22, and was last updated Nov 20, 2025.
Aider is the reason you've seen "Aider Polyglot leaderboard" cited in almost every serious coding-model comparison since 2024. It's free, Apache-2.0, works with any LLM you bring, and has 49,000+ GitHub stars to show for it. It's also had no tagged stable release since February 2026 and no commits to its main branch since May 22, 2026 — with about 1,400 open issues and about 500 open pull requests waiting (Oct 4, 2026). A community fork has quietly become the more actively developed branch of this family. This review is about what that actually means for you.
| Model flexibility (BYOK) | Excellent |
| Cost efficiency | Excellent |
| Maintenance velocity | Weak (since May 2026) |
| Bug-fix responsiveness | Weak |
| Benchmark credibility | Strong |
Aider still works. The question is whether "still works" is good enough in 2026.
"None of this makes Aider unusable... it installs cleanly and works with current models, but a bug you hit today may not get fixed soon."
That line, from independent reporting on Aider's 2026 status, is the fairest one-sentence summary I found. Aider is not abandoned — it's a working, free, genuinely useful tool with real engineering behind its core ideas (repo mapping, the diff-based edit format, the leaderboard methodology). But a project that used to ship every couple of weeks hasn't tagged a stable release since February and hasn't touched its main branch since May. That's not a rumor; it's what GitHub's own API returns when you ask it today.
A thin, git-native layer between you and whatever LLM you bring.
Aider is a terminal tool, not a platform: you run it in a repo, point it at an LLM (cloud API or local), and it edits your files directly, using git to auto-commit every change it makes so you can review or revert with normal git tools. There's no account, no subscription, and no vendor lock-in on the model — that architecture hasn't changed, and it's still the thing people like most about it.
The practical takeaway: the tool itself is unglamorous and reliable by design — it does one job (edit files via git) and leans on your own model choice for capability. What's changed isn't the design; it's how quickly anyone is fixing what breaks.
Aider's real contribution to the industry might be its leaderboard, not its user count.
Aider crossed roughly 49,000 GitHub stars and 6.8 million PyPI installs (per the project's own self-reported README badges, last regenerated around its most recent commit activity). Its "Aider Polyglot" benchmark — 225 hard Exercism exercises across 6 languages — is cited across the industry, including by competitors, as a credible measure of whether a model can actually edit real code, not just describe it.
Live GitHub count, checked Oct 4, 2026.
Project's own README badge, last generated ~May 2026.
Project's own README badge, last generated ~May 2026.
Project's own README badge, last generated ~May 2026.
| Signal | Aider-AI/aider (upstream) | cecli (community fork, formerly Aider CE) |
|---|---|---|
| Last stable release | v0.86.2, Feb 12, 2026 | v1.6.0, Sep 19, 2026 |
| Recent release cadence | None since Feb 2026 | 5 releases in the last month |
| Open issues | ~1,400 (Oct 4, 2026) | 10 |
| GitHub stars | 49,365 (Oct 4, 2026) | 407 |
| Extra features vs upstream | — | Agent mode, subagents, MCP support |
I'm not treating "6.8M installs" and "about 1,400 open issues" as contradictory — they're both true, describing a large, widely-installed tool whose maintenance pace has genuinely slowed, not a tool nobody uses.
The complaints aren't about code quality. They're about whether anyone's still driving.
I read Aider's own GitHub issues (not aggregated commentary about them, the actual threads) and independent 2026 coverage of its status, and kept the two separate.
"Thoughts and questions about the future direction of aider" — still open more than 8 months later, with only 2 comments and no maintainer reply. One commenter there describes the project as "essentially one man project."
"[Question] Is Aider stopped updating since four months ago?" — a user asking the exact question this review answers, directly on the project's own issue tracker, three weeks before this review was written.
Balanced independent assessment: Aider "installs cleanly and works with current models, but a bug you hit today may not get fixed soon" — usable today, riskier as a long-term dependency.
Aider itself is free. Your model bill is the real price tag.
Aider
Always free, Apache-2.0.
- No account, no subscription
- No usage limits from Aider itself
Local / Ollama
Run a local model.
- No API bill at all
- Weaker coding quality than frontier models
Frontier API key
Pay the model provider directly.
- Claude, GPT-5, Gemini, DeepSeek, etc.
- Pay only for tokens you use
OpenRouter
One key, many models.
- Aider ranks in OpenRouter's Top 20 apps (self-reported)
- Easy model-switching without new accounts
cecli (fork)
Same BYOK cost model.
- Same license, actively maintained
- Adds agent mode + MCP
Sources: pip/PyPI (Aider itself, free), individual model providers' own pricing pages (Anthropic, OpenAI, Google, DeepSeek, OpenRouter) for the actual token bill, verified Sep 21, 2026.
Your real decision is a model choice, not an Aider plan.
Pair Aider with whatever currently tops the Aider Polyglot leaderboard (GPT-5 (high) led at 88.0% as of this review).
Check the leaderboard's cost-per-task column, not just the accuracy column.
Point Aider at a local model via Ollama — quality will trail frontier APIs.
Check cecli's changelog before waiting on upstream — it may already be fixed there.
Upstream Aider doesn't have these yet — cecli does.
The leaderboard may be Aider's most important export, independent of the CLI itself.
Aider's polyglot benchmark runs 225 of Exercism's hardest coding exercises across C++, Go, Java, JavaScript, Python, and Rust, scoring "percent correct" as the share where all hidden unit tests pass after a model's second attempt. As of this review's check (Oct 4, 2026), GPT-5 (high) still tops the 69 models listed on the leaderboard at 88.0%, with o3-pro and various Claude-family models slightly behind it. The leaderboard itself was last updated November 20, 2025; its newest entry is dated October 3, 2025. This benchmark is referenced far beyond Aider's own user base — it's one of the standard reference points the whole industry uses to argue a model can actually edit code, not just talk about it.
| Metric | Value |
|---|---|
| Benchmark size | 225 exercises, 6 languages |
| Models listed (leaderboard last updated Nov 20, 2025) | 69 |
| Top score (Sep 2026 check) | GPT-5 (high), 88.0% |
| "Singularity" metric (self-reported) | 88% of new code in Aider's last release was written by Aider itself |
That last row is worth pausing on: Aider's own README reports that 88% of the new code in its last release was written by Aider working on itself — a genuinely interesting, self-referential data point, and also a reminder that a single maintainer using an AI pair programmer can still be a bottleneck if that one person's available time shrinks.
- The leaderboard methodology is genuinely useful independent of the CLI.
- Repo map + diff-based edits remain a clean, low-overhead architecture.
- True bring-your-own-model freedom, including fully local/offline.
- Apache-2.0 means you can fork or patch it yourself if you must.
- No agent mode, subagents, or native MCP support in upstream as of this review.
- Bug-fix latency is now genuinely uncertain.
- Self-reported install/usage badges are likely months stale.
- No official company, support channel, or roadmap commitment.
Aider has no company or official channel, so this is an independent tutorial, not official Aider content.
Git is the safety net the whole workflow depends on.
There's no in-app approval gate like a sandbox or permission prompt — the auto-commit-to-git pattern is the whole safety model. That's elegant when you're disciplined about reviewing diffs, and risky if you let commits pile up unread.
Watch mode is the closest thing Aider has to "automation" — deliberately limited.
Aider's "use in your IDE" watch mode monitors your files for specially-formatted AI comments and turns them into edit requests without you switching to the terminal. It's a lighter-weight automation model than an agent framework with subagents or scheduled background tasks — which is exactly the gap the cecli fork's "agent mode" is built to close.
- Single-file, well-scoped edits with clear intent
- Refactors where you want every step in a readable git log
- Local/offline work with a self-hosted model
- Quick prototyping where BYOK cost control matters
- Multi-step autonomous task chains (no native agent mode)
- Native MCP server connections (not in upstream)
- Long unattended background runs
- Team-shared configuration/governance features
MCP is the clearest capability gap between Aider and its own fork.
As of this review, upstream Aider does not ship native Model Context Protocol support. The community fork cecli does, alongside agent mode and subagents — which is precisely the feature set independent coverage cites as the reason serious users are migrating there rather than waiting on upstream.
| Capability | Aider (upstream) | cecli (fork) |
|---|---|---|
| MCP server support | Not available | Available |
| Agent mode / subagents | Not available | Available |
| Repo map, diff edits, git auto-commit | Available | Inherited from Aider |
| License | Apache-2.0 | Apache-2.0 |
There is no "Aider bill" — only your model bill.
Because Aider is a thin client over whatever LLM you choose, its real economics are just that model's own API pricing. The leaderboard's cost column (tracked alongside accuracy) is the most honest way to compare "value," not Aider's own pricing page — because Aider doesn't have one.
| What you're doing | What you pay for |
|---|---|
| Running Aider itself | Nothing — free, Apache-2.0, no account |
| Using a frontier model (Claude, GPT-5, Gemini) | That provider's own per-token API pricing |
| Using OpenRouter as an aggregator | OpenRouter's per-model pass-through pricing plus its own margin |
| Using a local model via Ollama | Your own hardware/electricity only |
No sandbox, no approval prompts — git is the entire control surface.
Aider does not run its own sandbox or approval-mode system the way Codex or Gemini CLI do. It edits files directly in your working directory and relies on git auto-commits as the review/undo mechanism; any shell commands it suggests (for linting or testing) still go through your terminal, under your normal user permissions. This is a materially different trust model than a sandboxed agent CLI, and worth knowing before you point it at a sensitive repo.
| Control | How Aider handles it |
|---|---|
| File edit review | Git auto-commit per change; review/revert via normal git tools |
| Shell command execution | Runs under your own user permissions; no built-in sandbox |
| Credential/API key handling | You supply your own provider API key directly; no Aider-hosted account |
| Data retention | Governed entirely by whichever model provider you choose, not by Aider |
For anything genuinely sensitive, I'd treat Aider the way I'd treat any script you're running locally with your own credentials: know what it's about to touch before you let it commit.
This is a tool for developers who want control, not hand-holding.
| Work pattern | Why Aider fits | Watch-out |
|---|---|---|
| Developer who wants model choice, no vendor lock-in | True BYOK across every major provider | None specific to this use case |
| Cost-conscious solo developer | Pay only for tokens, no seat fee | None specific |
| Team wanting agent mode / MCP today | Not upstream Aider's current strength | Consider cecli instead |
| Team needing guaranteed support/SLA | No company stands behind this project | Real risk if you depend on fast fixes |
| Privacy-sensitive/offline work | Full local-model support via Ollama | Quality trade-off vs frontier APIs |
This is the section where I stop being diplomatic about the maintenance question.
v0.86.2 (Feb 12, 2026) is the last tagged, published version on PyPI as of this review — over seven months with nothing new shipped to users who don't build from source.
Live GitHub data shows the last push to the default branch was in late May — roughly four months of silence as of this review, confirmed via the repository's own API.
A backlog that size, with a maintainer described in the project's own issue tracker as running it as "essentially one man project," is a real risk if you hit a bug.
Agent mode, subagents, and MCP support exist in cecli but not in upstream Aider — the ecosystem has effectively already voted on where active development is happening.
If Aider's pace is the problem, here's where else to look.
| If you mostly need... | I'd compare Aider with... |
|---|---|
| The actively-developed branch of this exact project family | cecli (github.com/cecli-dev/cecli) |
| A vendor-hosted terminal agent bundled with a paid subscription | OpenAI Codex — read our Codex review |
| Google's free-tier-friendly open-source terminal agent | Gemini CLI — read our Gemini CLI review |
| A terminal agent bundled with a Claude subscription | Claude Code — read our Claude Code guide |
| A polished, opinionated standalone editor | Cursor (cursor.com) |
Nobody paid for a kinder verdict, including Aider itself.
At the time of this review, JAVIS has not established a publisher affiliate program with Aider or its maintainers — there is no company to have one with. Every CTA in this article points to the official open-source repository.
Go to the official repository.
Open AiderRead the alternative that matches your real question.
Choose the next articleCoding-agent reviews only make sense read next to each other.
Aider, Codex, and Gemini CLI are worth reading together because they represent three different bets on how much you should pay, and to whom, for an AI pair programmer. These are the pages I'd read next.
Questions worth answering before you rely on this
Is Aider still maintained?
The GitHub repository is still online and the license hasn't changed, but live data shows no stable release since Feb 12, 2026 and no commits to the main branch since May 22, 2026, as of this review. It's not deleted or archived, but active development has visibly slowed.
Is Aider free?
Yes, Aider itself is free and Apache-2.0 licensed. Your cost is whatever LLM API you connect it to.
What is cecli / Aider CE?
A community fork of Aider (Apache-2.0, same lineage) that has become the more actively maintained branch, shipping multiple releases in September 2026 alone and adding agent mode, subagents, and MCP support that upstream Aider does not have.
Does Aider support MCP?
Not in the upstream Aider-AI/aider project as of this review. The cecli fork does.
What model should I use with Aider?
Check the current Aider Polyglot leaderboard rather than assuming — as of this review GPT-5 (high) led at 88.0%, but rankings shift as new models ship.
Is Aider safe to run on a real codebase?
It has no built-in sandbox; it relies on git auto-commits as its review/undo mechanism. Treat it like any local script running with your own credentials, and review diffs before you trust them.
Should I switch to cecli instead of upstream Aider?
If you specifically need agent mode, subagents, or MCP, or you've hit a bug that isn't getting fixed, it's worth trying — same license, same core lineage, currently much faster-moving.
Who maintains Aider?
It was built and has been primarily maintained by Paul Gauthier under the Aider-AI GitHub organization; community commentary in the project's own issue tracker describes it as effectively a one-person effort.
Where this came from, and when it was checked.
Repository stats, license, last commit/push date: GitHub data for github.com/Aider-AI/aider, Sep 21, 2026.
Last stable release version and date: pypi.org/project/aider-chat, official PyPI package metadata, Sep 21, 2026.
Community-concern evidence (Issues #4751, #5673): the GitHub issue threads themselves, confirming titles, open state, comment counts, and dates directly, Sep 21, 2026.
cecli / Aider CE fork data: GitHub data for github.com/cecli-dev/cecli (stars, forks, open issues, release history), Sep 21, 2026.
Product README, features, official images: raw.githubusercontent.com/Aider-AI/aider/main/README.md and aider.chat, read directly, Sep 21, 2026.
Aider Polyglot leaderboard (GPT-5 (high), 88.0%, 69 models listed; last updated Nov 20, 2025): aider.chat/docs/leaderboards/ and independent benchmark-tracking sites, cross-checked Oct 4, 2026.
Self-reported install/usage badges (6.8M installs, 15B tokens/week, Top 20 OpenRouter, 88% Singularity): quoted directly from the official README's auto-generated badge text, explicitly labeled as self-reported and likely last-regenerated around the project's most recent commit activity.
Independent video: youtube.com/@Snyksec, "Can You Really Pair Program with AI? (Aider & Cohere)," authorship confirmed on YouTube; explicitly labeled as independent, not official Aider content.
