The best AI coding agents in November 2026 are Claude Code and Codex, which hold every top spot on the public Terminal-Bench 4.0 leaderboard, followed by OpenCode, the strongest open-source agent with free choice of model. GitHub Copilot is the best free start and the easiest pick for teams on GitHub, and Cursor is the best AI editor. This ranking was last updated on 4 November 2026; the field moves monthly, so check the date.
Every placement below follows the criteria in the next section.
How this ranking works
Five criteria, in this order of weight:
- Measured results. How the agent scores on a public benchmark that runs real tasks end to end. Only Terminal-Bench 4.0 ranks agents rather than just models, so it carries the most weight, with its limits (below).
- Model choice. Can you use more than one vendor’s models, and local ones? An agent tied to one vendor inherits that vendor’s prices, limits and outages.
- Where it runs. Terminal, editor, desktop app, cloud. And whether it works natively on Windows, which many developers use and many agents treat as an afterthought.
- Control. How you decide what the agent may do without asking: rules you can read, a reviewer model, or nothing.
- Cost model and maintenance. How you pay (subscription, per token, free tier) and whether the project still ships. Prices change too often to compare as numbers; each vendor’s pricing page is linked.
What the benchmark says
Terminal-Bench 4.0 runs each agent and model pair on hard terminal tasks and reports how many it solves. The top rows on 4 November 2026:
| Agent | Model | Reasoning effort | Tasks solved |
|---|---|---|---|
| Claude Code | Opus 5.5 | max | 64.85% |
| Claude Code | Sonnet 5.5 | max | 61.82% |
| Codex | GPT-6 Astra | max | 58.18% |
| Codex | GPT-6.1 Sol | max | 58.18% |
| Claude Code | Fable 5.1 | max | 57.88% |
| Claude Code | GLM-5.3 | max | 41.82% |
| Grok Build | Grok 4.7 | xhigh | 37.58% |
Three things the table shows that a ranking list usually hides:
- The gaps are smaller than they look. Each score carries a margin of about plus or minus 3 points, so the rows from 58% to 62% overlap.
- The model matters more than the agent. The same Claude Code scores 64.85% with Opus 5.5 and 41.82% with GLM-5.3.
- Open-source agents aren’t on it. OpenCode, Cline, Kilo Code and the rest have no submitted rows, so the benchmark can’t rank them. Their placement below rests on criteria 2 to 5.
For how these benchmarks work and how far to trust them, see AI coding benchmarks.
The ranking
| # | Agent | Where it runs | Models | Open source | Windows | How you pay |
|---|---|---|---|---|---|---|
| 1 | Claude Code | Terminal, VS Code, JetBrains, desktop, web | Claude only | No | Native; sandbox only in WSL 2 | Claude plan or per token |
| 2 | Codex | CLI, IDE extension, ChatGPT desktop app, cloud | OpenAI; local via Ollama or LM Studio; providers that speak the Responses API | CLI: Apache-2.0 | Native, with its own Windows sandbox | ChatGPT plan (Free included) or per token |
| 3 | OpenCode | Terminal, desktop app (beta), IDE extension | 75+ providers, ChatGPT and Copilot sign-in, local | MIT | Native; docs recommend WSL | Your provider; optional OpenCode Go plan |
| 4 | GitHub Copilot | VS Code and other editors, CLI, Copilot app, cloud agent | Many vendors through GitHub | No (CLI) | Native | Copilot plan, Free tier |
| 5 | Cursor | Its own editor, CLI | Many vendors through Cursor | No | Native | Cursor plan, Free tier |
| 6 | Cline | VS Code, JetBrains, CLI, desktop app | Many providers, your keys | Apache-2.0 | Native | Your provider |
| 7 | Kilo Code | VS Code, JetBrains, CLI, desktop app (beta), cloud | 500+ models via Kilo Gateway, your keys, local | MIT | Native | Your provider or Kilo credits |
| 8 | goose | Desktop app, CLI | 15+ providers, local | Apache-2.0 | Native desktop app | Your provider |
| 9 | Pi | Terminal, SDK | Many providers | MIT | Native installer | Your provider |
| 10 | Aider | Terminal | Most cloud and local models | Apache-2.0 | Via Python | Your provider |
| 11 | DeepSeek Harness | Local web UI | DeepSeek, other providers, OpenAI-compatible endpoints | MIT | Not stated | Your provider |
| 12 | Antigravity | CLI (agy), IDE, 2.0 desktop app, editor extensions | Gemini, plus a fixed list that includes Claude models | No | Native, with a sandbox on Windows 11 24H2 | Google AI plan (Free, Pro, Ultra) or an API key |
| 13 | IBM Bob | Bob IDE, Bob-Shell CLI, IBM cloud or on-premises | Claude, Mistral and IBM Granite, routed per task | No | Native | Per instance, plus paid add-ons |
| - | Contenox | Desktop app | ChatGPT and Claude (through Claude Code) plans, Copilot as an unofficial sign-in; API keys for many providers; local | Apache-2.0 | Windows first; sandbox limits writes to the project | Your provider |
Why each one sits where it does:
- Claude Code has the top two benchmark rows and runs everywhere you work. It loses points only on model choice: Claude only, and no free plan.
- Codex is a few points behind Claude Code on the benchmark, level with Claude Code on Fable 5.1, and does more on Windows: it runs natively in PowerShell with its own sandbox, where Claude Code’s sandbox needs WSL 2. Its CLI is open source, and ChatGPT’s Free plan includes it.
- OpenCode is the best agent you can run with any model. It signs in with a ChatGPT or Copilot plan, takes API keys for 75+ providers and talks to Ollama, LM Studio and llama.cpp.
- GitHub Copilot is the widest product: editor, terminal, app, cloud agent and code review inside GitHub, with a Free plan that includes agent mode and the CLI.
- Cursor is the most polished AI editor, with the best tab completion. Its agent is strong, but you work inside Cursor and through Cursor’s servers.
- Cline brings an agent into VS Code or JetBrains with any model and your own keys.
- Kilo Code, owned by Anaconda since July 2026, covers the same ground as Cline, with its own gateway to hundreds of models and a free option that routes to free models.
- goose is a general-purpose agent, not just for code, now under the Agentic AI Foundation at the Linux Foundation.
- Pi is a small agent you extend yourself. It has no permission prompts by design, so run it in a container.
- Aider pioneered the terminal pair programmer, but its last release on PyPI was in February 2026.
- DeepSeek Harness is promising and widely starred, but it is a developer preview that warns of breaking changes.
- Antigravity is Google’s agent, three ways: the
agyCLI that replaced Gemini CLI for personal accounts in June 2026, the Antigravity IDE, and the 2.0 desktop app from I/O 2026 that runs several agents in parallel. It has no benchmark row of its own and no open-source license, and the IDE is not supported for enterprise accounts, which is why it sits here rather than next to Claude Code and Codex. See Google Antigravity. - IBM Bob is the enterprise one: generally available since April 2026, on-premises and air-gapped since October, with a model router instead of a model picker. It sits last because it has no benchmark row and its price is per instance with add-ons for Java, IBM i, Z and DevSecOps. If your code can’t leave the building, it moves to the top. See IBM Bob.
Best AI coding agent for each need
| Need | Pick | Why |
|---|---|---|
| Best overall | Claude Code | Top benchmark rows, every surface. Codex is a close second. |
| Best free | GitHub Copilot Free | Agent mode and the CLI are included, with limited usage. Codex in ChatGPT Free and Cursor’s Free plan are the alternatives. |
| Free with no limits | OpenCode with a local model | No subscription and no per-token bill; your hardware sets the speed. See best Ollama models for coding. |
| Best on Windows | Codex | Native PowerShell with its own sandbox. See AI coding agents on Windows. |
| Best for local models | OpenCode | Ollama, LM Studio and llama.cpp as first-class providers. |
| Best open source | OpenCode | MIT, any provider, terminal and desktop. See open source AI coding agents. |
| Best inside an IDE | Cursor | The best editor experience. To stay in VS Code or JetBrains, Cline or Copilot. |
| Best for teams | GitHub Copilot | Organization plans, policies, the cloud agent and code review where your code already lives. |
| Best on-premises | IBM Bob | Runs inside your own data center or air-gapped, with Claude, Mistral and Granite behind one router. |
The cost gotcha: “free” agents still send your whole context to a paid model on every step. An 8-prompt agentic session can send about 300 million input tokens. The Contenox team measured one such session at about $80 on Claude Fable 5.1 with all optimizations, and under $1 on DeepSeek V4 Flash (since replaced by V4.1 Flash). The agent is rarely the expensive part; the model is. See AI coding subscriptions compared.
Which agents dropped off older lists?
Many “best AI coding tools” lists and Reddit threads still recommend tools that changed in 2026:
- Roo Code was archived on 15 May 2026. The repository is read-only and the extension gets no more fixes. Kilo Code and Cline are the closest replacements.
- Continue is no longer actively maintained. Its final 2.0.0 release shipped, and the repository is read-only.
- Gemini CLI no longer serves free or Google AI Pro and Ultra accounts. Since 18 June 2026 it is for Gemini Code Assist Standard and Enterprise and paid API keys; Google points everyone else to Antigravity CLI.
- Void, the open-source AI editor, is deprecated.
Contenox
Contenox isn’t ranked: it isn’t on the benchmark, and it’s a younger project than the tools above. It fits one need: a Windows desktop app where you open a project folder, choose any provider (your ChatGPT or Claude plan, an API key, or a local model) and approve the agent’s work against permission rules you can read. Out of the box it reads freely, runs builds and tests without asking, and asks before edits and risky commands. On Windows, a sandbox limits writes to the project folder. It is free and open source; download it here.
Related guides
- AI coding: the three kinds of tools, and which fits which job
- IBM Bob: the enterprise agent, inside the firewall
- What is an AI coding harness?: why the same model scores differently in different agents
- AI coding harness comparison: features side by side
- Terminal coding agents: how to install the top three
- AI coding benchmarks: what the scores measure
- Claude Code alternative and Codex alternative