It was eleven on a Friday night when Chen stared at a refactor spanning a dozen files and did the math. The terminal AI coding assistant he relied on cost him more than a hundred dollars a month, yet the project only really needed it two or three hours a day. He started digging through GitHub for an open-source replacement, and found the field far more crowded than he expected.
This piece starts from Chen’s problem. It is not a comprehensive benchmark. The question is narrow: if you want an AI coding assistant that lives in your terminal, what are the solid choices in 2026, and who is each one for?
Meet the headliner: where Qwen Code comes from
Qwen Code is the open-source terminal AI coding assistant from Alibaba’s Tongyi team. Open its GitHub repo and you will spot a modest line in the docs: it is forked from Google’s Gemini CLI. That is no secret, and it is written plainly in the project. Standing on someone else’s work is routine in open source. What matters is what you build on top.
Qwen Code has done plenty of its own building. It is tuned deeply around the Qwen model family, especially Qwen3-Coder, while keeping the ability to plug in other models. OpenAI, Anthropic Claude, Google Gemini, even a local model through Ollama, all of them can be wired up. On features, it has added SubAgents, Agent Teams, cross-session memory, MCP tool integration, and a Skills system, the kind of thing heavy users ask for. By September 2026 it had gathered close to 28,000 stars on GitHub, with an active community and weekly releases.
One change is worth flagging. Qwen Code used to offer a free OAuth login quota, so signing up for a Tongyi account got you going right away. That free tier was removed in April 2026. To use Qwen Code now, you bring your own API key: a Model Studio (Bailian) plan from Alibaba Cloud, a third-party model key, or a local open-source model.
Which raises the obvious question. If you are supplying the model anyway, why not look at the other tools?
Claude Code: the reference experience, at a price
Ask on social media which terminal AI coding assistant is best, and someone will say Claude Code.
Anthropic polished this terminal agent well. It reads code, edits files, runs commands, and handles Git operations, with a context window that stretches to a million tokens, so it rarely loses the thread on a large codebase. Its agent behavior is mature too. Faced with a complex task it breaks the work into steps, and when a test fails it tries to fix the problem itself. That self-correction leads the pack.
The catch is cost. Claude Code is not sold on its own. It rides inside Claude’s subscription tiers. The entry point is the Pro plan at 20 dollars a month, which covers maybe an hour or two of serious use per day. Heavy users move up to the Max plan at 100 or 200 dollars a month, depending on whether you want the 5x or 20x quota. The other route is pay-as-you-go API billing, with no monthly cap and no ceiling either.
There is a deeper issue. Claude Code is closed source and locks you to Claude models. You cannot point it at GPT, Gemini, or any local model. For an individual developer that may just be money. For a company, closed source plus model lock-in means your development workflow leans hard on one vendor, which carries risk on both compliance and bargaining power.
Chen started on Claude Code’s Pro plan. The experience was good, but he kept hitting the quota limit every afternoon and had to sit and wait for hours. Upgrading to Max felt expensive. That is how his Friday night began.
Gemini CLI: the upstream, and Google’s open-source play
Since Qwen Code forks from Gemini CLI, the original deserves a look.
Gemini CLI is the terminal AI agent Google open-sourced in 2025 under the Apache 2.0 license, with more than 100,000 stars on GitHub. It ties into Google’s Gemini models, though as an open project the community has adapted it to reach other models. Google offers a decent free quota. Sign in with a Google account and an individual developer can use it for free, with a daily request cap that covers light and moderate use. Need more, and you can buy a Google AI paid plan.
Gemini CLI shines when you are already inside Google’s ecosystem. If you use Google Cloud, or your project runs on GCP, it brings some natural convenience. Step outside that ecosystem and the advantage disappears.
Here is a twist. Gemini CLI went through a rebrand in June 2026, when Google introduced Antigravity CLI, whose command is agy, as the new name. The old gemini command still works, and the community mostly keeps calling it Gemini CLI.
The relationship between Qwen Code and Gemini CLI resembles that of an Android phone maker and AOSP. They share the base, while the features, tuning, and ecosystem direction go their own ways up top. If you mainly use Qwen models or want strong support for Chinese development scenarios, Qwen Code fits better. If you value Google’s model capabilities and free quota, Gemini CLI is the more direct choice.
OpenAI Codex CLI: big-vendor open source with real sandboxing
OpenAI’s Codex CLI is another heavyweight. It has passed 120,000 stars on GitHub as of September 2026, is written mostly in Rust, and is open-sourced under Apache 2.0.
Codex CLI has one design highlight: it takes sandbox isolation seriously. You can set different security levels so the agent runs commands in a restricted environment, which keeps it from deleting files by accident or causing surprises. For developers working near production code, that detail earns a lot of goodwill.
On models, Codex CLI defaults to OpenAI’s own lineup. The current workhorse is the GPT-5.6 series, split into Sol, Terra, and Luna tiers for different performance and price points. Billing is flexible. You can spend a ChatGPT paid subscription quota (Plus starts at 20 dollars a month) or use an OpenAI API key on metered billing. New accounts get a small free quota to try.
Like Claude Code, Codex CLI stays inside OpenAI’s model family. It is open source, so the community could in theory adapt it to other models, but OpenAI never set out to make it model-agnostic.
Aider: the veteran that answers to no single model
Where the tools above carry some “our model first” DNA, Aider takes a different path entirely.
Aider is a pure open-source terminal AI pair-programming tool, written in Python, licensed Apache 2.0, with close to 50,000 stars on GitHub. Its defining trait is being model-agnostic. Claude, GPT, Gemini, DeepSeek, a local model through Ollama, Aider supports more than 70 model providers, and you switch with a single command. API down? Swap it. One vendor too pricey? Try another. That flexibility is hard to find elsewhere.
Aider’s other signature is deep Git integration. Every change it makes becomes an automatic Git commit with a clear message. So you can diff, roll back, or rebase at any point, and every edit the AI makes leaves a trail. For teams that care about code review, this is handy.
It also has an Architect Mode, where an expensive model handles planning and analysis while a cheaper model executes the actual edits. In practice that division cuts high-end model API costs by 30 to 50 percent, which appeals to cost-conscious developers.
Aider has no subscription fee. It is a pure tool: you bring your own API key and settle model costs directly with the provider. The upside is full transparency. The downside is that you manage the keys yourself and need to understand each vendor’s pricing, which is a small hurdle for beginners.
The IDE alternatives: Cline and Continue
Not everyone wants to chat with an AI in a bare terminal. If you are more at home in VS Code or JetBrains, IDE-plugin tools like Cline and Continue are worth a glance.
Cline is a VS Code plugin, open source, supporting multiple models, letting you talk to the AI, edit files, and run commands right in the editor. Continue follows a similar line, offering chat, autocomplete, and code editing inside the IDE, plus a CLI component that runs in a terminal or a CI/CD pipeline.
Their core difference from pure terminal tools is this. IDE plugins tap the editor’s existing syntax highlighting, code navigation, and debugger, so they feel smoother to pick up, while terminal tools tend to be lighter and easier to fold into automation scripts and headless environments. The two are not mutually exclusive. Many developers keep both an IDE plugin and a terminal agent installed and switch by scenario.
Side by side: one table for the core differences
Having walked through each tool, making a real choice still calls for laying the key dimensions out together. The table below lists only the dimensions that move a decision: how open it is, model lock-in, cost structure, and how it runs. Finer feature differences (whether it supports MCP, whether it has subagents) shift with every release, so the table leaves them out.
| Tool | Open source / license | Model binding | Cost | Form factor |
|---|---|---|---|---|
| Qwen Code | Open source, Apache-family license | Qwen by default, third-party and local models configurable | Free CLI, bring your own API key (Alibaba Cloud Bailian plan or third party) | Terminal CLI, plus IDE plugin and desktop app |
| Claude Code | Closed source | Claude models only | Pro from $20/mo, Max $100-200/mo, or metered API | Terminal CLI |
| Gemini CLI | Open source, Apache 2.0 | Gemini by default | Free Google-account quota, paid plans to scale up | Terminal CLI |
| Codex CLI | Open source, Apache 2.0 | OpenAI by default | ChatGPT subscription quota or metered API | Terminal CLI |
| Aider | Open source, Apache 2.0 | No lock-in, 75+ providers | Free tool, model costs on you | Terminal CLI |
| Cline / Continue | Open source | Multiple models configurable | Free tool, model costs on you | IDE plugin (VS Code / JetBrains) |
A table helps you filter fast, but selection cannot rest on the table alone. Here is how to choose across a few common real-world situations.
How to choose: who you are, and what you care about
Cost-sensitive solo developers. You might be building your own project or taking on freelance work, with a tight monthly tool budget. Here Aider is close to the optimal answer, since the tool itself is free and you only need a cheap API key (DeepSeek, for instance, runs far below the major vendors), and Architect Mode saves more still. Qwen Code is also a good pick, especially if you already have Alibaba Cloud resources, since the Bailian plan’s domestic pricing beats overseas APIs by a fair margin.
Teams already all-in on one cloud ecosystem. If your company leans hard on Alibaba Cloud, pointing Qwen Code at Bailian models is the path of least resistance. If you are on Google Cloud, Gemini CLI runs smoother. If OpenAI is your main model vendor, Codex CLI integrates best. Picking the tool that lines up with your existing infrastructure saves a lot of fiddling.
Enterprises that prize data security and self-control. Open source is non-negotiable here. Claude Code is out from the start, because closed source means you cannot audit the code, deploy internally, or run it offline. Among the rest, Qwen Code, Aider, and Codex CLI all carry Apache-family licenses, so you can modify and self-host freely. Add a local model (Qwen3-Coder or Code Llama through Ollama, say) and you can stand up an internal coding assistant that never touches an external API.
People who want the strongest out-of-the-box experience. If budget is no object, Claude Code’s Max plan leads on overall experience at this point in time, with mature agent behavior, a large context window, and high task-completion rates. Codex CLI paired with GPT-5.6 Sol sits in the first tier too. Both share a trait: they run their own best model, tuned the deepest, so they feel more fluid than model-agnostic tools. The price is model lock-in and an ongoing subscription bill.
Chinese development scenarios and domestic networks. This is a quiet advantage for Qwen Code. Its grasp of Chinese leans on Qwen’s Chinese training data, so it usually handles Chinese comments, docs, and requirement descriptions more naturally than overseas models. Add that Alibaba Cloud’s API needs no VPN from within China and has low latency, and daily use runs noticeably smoother for developers working there.
A trend that is easy to miss
One pattern stands out in the 2026 terminal AI coding tool space: open-source projects are learning from each other, and their features are converging fast. Qwen Code forked from Gemini CLI and then added Claude-Code-style subagents and autonomous coding modes. Codex CLI leads on sandbox security, yet Gemini CLI and Qwen Code are following with similar mechanisms. Aider’s Architect Mode has been borrowed by many tools.
So the differences at the pure feature level will keep shrinking. What decides your long-term choice tends to be the things that do not change easily: the license (can you modify it, can you self-host), model flexibility (will one vendor lock you in), and where you sit in the ecosystem (what cloud, what model, what workflow your team already runs).
What did Chen pick in the end? He tried the whole lineup, wavered a while between Aider and Qwen Code, and installed both. Day to day he runs Aider against the DeepSeek API for most tasks, and switches to Qwen Code for projects that need deep reading of Chinese requirement docs. Together, the two cost him less than half of his old Claude Code Pro subscription a month.
The terminal AI coding assistant space is still moving fast. No single tool solves everything for good. The good news is that the open-source options keep growing and maturing, and developers hold more choice than before. Finding the one that fits your current situation and putting it to work beats agonizing over the decision.
Related reading
- Claude Code vs Codex CLI vs Gemini CLI: Which Terminal AI Coding Assistant Feels Like a Teammate in 2026
- The Day AI Learned to Run Your Computer and the Open-Source Fortress Changed Hands
- Cursor vs Copilot vs Claude Code vs Windsurf: Which AI Coding Tool Is Right for You in 2026?
- Cursor 3 and the Shift from Writing Code to Managing Agents
- Which AI Email Assistant Actually Saves You Time? Superhuman AI vs Spark AI vs Notion Mail vs Gmail AI in 2026
Stay updated with our latest AI insights
