Best new AI coding agents (October 2026)

Overall best AI coding agent

Short answer: our overall pick is Claude Code, an editorial #1 rather than a benchmark result. In our smoke test on 2026-10-25 it partly worked with a small local model (not Claude models), so the test says little about how it performs with its vendor's own models.

• Best overall (terminal): Claude Code, our editorial #1. It needs a paid Claude plan and is not on the Free plan.
• Best for IDE-based daily work: Cursor and GitHub Copilot remain practical choices.
• Best open-source, model-flexible terminal agents: OpenCode, Cline, Goose and Qwen Code. OpenCode and Goose each partly worked in our smoke tests on a small local model.
• Which model to run inside an agent: see best LLM for coding for sourced benchmark numbers with dates.

AI coding agents compared: price, free tier and best use

Prices change and several vendors bill by usage or credits, so a monthly price is not the whole cost. "Not verified" means we did not find the figure on an official page we read, not that none exists. Tools marked "not yet a full entry" have no Agentic Weekly write-up yet; their rows come from the linked official pages only and nothing about them was run. Evidence labels: see methodology.

Best picks by use case

These are editorial picks based on what each vendor says and, for Tested entries, the small smoke test we ran. They are not benchmark results.

• Terminal-first, most actively updated: Claude Code, our #1. It needs a paid Claude plan (it is not on Free) and our test used a local model, not Claude models. See Claude Code vs Cursor.
• You live in an editor: Cursor, or GitHub Copilot if your team already works in GitHub. Copilot's Free plan supports CLI and agent mode with limited usage, per its plans page.
• You want open-source software you can read: Codex CLI, OpenCode, Cline, Goose and Qwen Code are all open source; you pay for whichever model you connect.
• You want to switch model providers freely: OpenCode, Cline and Qwen Code advertise support for many providers; Qwen Code lists OpenAI, Anthropic, Gemini and Qwen APIs plus local models.
• You want agents to keep working while you are away: Amp runs each thread in a remote machine it calls an orb, and its docs say the agent keeps working after you close your laptop.
• You are on Google's stack: Antigravity CLI is Google's terminal agent. Gemini CLI's README says unpaid-tier and Google One users moved to Antigravity CLI on 2026-06-18, so check which one applies to your account.
• You want to spend nothing to try one: Cursor and GitHub Copilot both list free plans with limits; Amp has a free Hobby tier; the open-source tools cost nothing but still need a model.

New and notable coding agents we have read about but not yet written up

These are real products with official documentation we read on 2026-10-25. We have not run any of them, and we do not rank them. Each will get a full entry only when we can say what we tested.

• [Antigravity CLI](https://antigravity.google/docs/cli/overview/) (agy): Google's keyboard-driven terminal agent. Its docs say it shares an agent harness and settings with Antigravity 2.0 and installs with a one-line script on macOS and Linux. It also runs non-interactively with agy -p.
• [GitHub Copilot CLI](https://docs.github.com/en/copilot/get-started/cli-quickstart): installs with npm install -g @github/copilot (Node.js 22 or later) and requires an active Copilot subscription unless you bring your own model provider keys, per GitHub's docs. It is part of the Copilot entry above.
• [Amp](https://ampcode.com/docs): a coding agent and development environment that runs each thread in a remote orb. It works from the web, a terminal CLI and mobile apps, and its pricing page lists a free Hobby tier and a $20/month Individual tier.
• [Qwen Code](https://github.com/QwenLM/qwen-code): an Apache-2.0 terminal agent from the Qwen team. It installs with npm or Homebrew and you run /auth to configure a provider.
• [Kimi Code](https://www.kimi.com/code/docs/en/): Moonshot AI's coding service, offered with a Kimi membership as a desktop app, a CLI and a VS Code extension. Its docs say the CLI was rebuilt on Node.js.
• [Devin CLI](https://docs.devin.ai/cli): Cognition's local terminal agent with optional hand-off to Devin Cloud, installed with a one-line script.
• [Pi](https://github.com/badlogic/pi-mono/blob/main/packages/coding-agent/README.md): a deliberately minimal terminal agent whose README describes interactive, print and JSON, RPC and SDK modes and TypeScript extensions.

How to choose a coding agent

Start with where you work. If you spend your day in a terminal, the terminal agents (Claude Code, Codex CLI, OpenCode, Goose, Qwen Code, Antigravity CLI) fit naturally. If you want the agent next to your files, an editor product such as Cursor, or an extension such as Cline or Copilot, is a smaller change.

Next, decide who pays for the model. Claude Code, Cursor, Copilot and Amp bundle model usage into a plan, so you pay the vendor and accept its limits. The open-source agents are free to install but send requests to a model you pay for, or run a local one. Local models are cheaper but, in our own smoke tests with a small local model, were far weaker than the hosted models these tools are designed for.

Finally, check how fast the tool changes. These agents release almost daily, so a pick today may be stale next month. Each tool page shows the release we covered and the date we last verified it, and the weekly archive shows what changed.

How these rankings work

The ranked list above is editorial judgement on the tools we cover. It is not a benchmark, a popularity count or a paid position. Paid placements are labelled and never change a rank. The full basis is on the methodology page. A Tested label means a small smoke test with its scope stated: our tests ran on a Linux machine with no GPU, using a small local model, so they say little about how a tool performs with its vendor's own models.

Side by side: what we recorded for the top 8

Every cell is read from the tool’s own pages on the date shown on its entry. “Not recorded” means we have not verified it, not that it does not exist.

Who each pick suits

• Claude Code: Good for: Anyone relying on the Bash sandbox or permission rules (the 2.1.309 fixes close real gaps); Teams running Claude Code through Bedrock, Vertex or Foundry; Enterprise admins who want role-based control over cloud sessions and Remote Control. Skip it if: You are looking for new headline capabilities; these two releases are mostly fixes; You script claude ultrareview; check the new --trust-workspace requirement before updating CI.
• Codex CLI: Good for: People who already use Codex and repeat the same setup prompts (presets save the typing); Teams that connect Codex to enterprise MCP servers and want fewer sign-in hurdles; Anyone who wants an open-source (Apache-2.0) agent they can run from a terminal. Skip it if: You are happy with another coding agent; nothing here changes the core capability; You need the image preview feature but your terminal does not support Kitty or Sixel graphics.
• Gemini CLI: Good for: Organizations on Gemini Code Assist Standard or Enterprise (Google says their access is unchanged); People using Gemini CLI with a paid Gemini API key; Developers who want an open-source Apache-2.0 terminal agent they can read and extend. Skip it if: You are on a Google AI Pro/Ultra plan or the free individual tier: Google says Gemini CLI stopped serving those requests on June 18, 2026; You want the best coding experience with the latest Gemini models: the README itself points to Antigravity CLI.
• Cursor: Good for: Teams that want cloud agents with shared, scriptable environments; Developers who like an editor-first workflow with an agent built in; Platform teams evaluating Origin and its app marketplace. Skip it if: You work only in a terminal and want a lightweight open-source agent; You want all third-party model usage included in a flat fee; Cursor says it is billed separately by usage.
• GitHub Copilot: Good for: Teams already on GitHub who want AI in editor, CLI and code scanning fixes; Enterprise admins who need model policy controls; Developers who want to pick from several vendors' models. Skip it if: You want a fully open-source stack; You are on a plan without access to the newest models (Gemini 4 Argon is Pro+ and above).
• opencode: Good for: Developers who want an MIT-licensed terminal agent; People who switch between model providers; Anyone who wants a desktop app as well as a CLI. Skip it if: You need a vendor-backed support contract; You expected feature news; this release is mostly a runtime upgrade.
• Cline: Good for: IDE users who want an open-source agent; Teams that want to embed an agent through the SDK; People who bring their own model API keys. Skip it if: You want one polished all-in-one app with bundled models; You only want headline features; these releases are mostly fixes and plumbing.
• goose: Good for: People who want an Apache-2.0 agent with a desktop app and CLI; Users who delegate work to subagents and need resume and limits; Windows desktop users who were waiting for an installer. Skip it if: You want a tiny tool with few moving parts; You need a managed, supported product.

Compare the options

• Codex CLI vs Gemini CLI: which should you use?
• Augment Code vs Devin: which should you use?
• Claude Code vs Codex CLI: which should you use?
• Claude Code vs Cursor: which should you use?
• Claude Code vs Gemini CLI: which should you use?
• Cursor vs GitHub Copilot: which should you use?
• Cursor vs Windsurf (now Devin Desktop): which should you use?
• GitHub Copilot vs Qodo: which should you use?
• goose vs opencode: which should you use?
• Kiro vs Amazon Q Developer: which should you use?
• OpenHands vs Devin: which should you use?
• Qodo vs Amazon Q Developer: which should you use?
• Zed vs Kilo Code: which should you use?

Looking for something else? See all best-for categories, or the quick-answer tables.

Setup guides

• How to add an MCP server to Claude Code
• Best model for OpenCode, Cline and Claude Code: options by budget
• ChatGPT Work vs ChatGPT Chat vs Codex: what each is for
• Is Claude Opus 5.5 free? Which plans include it
• GitHub MCP server: how to set it up
• Which coding agents can use GPT-6.1 Terra?
• How to run Claude Code with a local model
• Which Claude model should you use? Haiku 5.5, Sonnet 5.5, Opus 5.5 or Fable 5.1

FAQ

Sources

• Claude pricing (read Oct 25, 2026)
• Cursor pricing (read Oct 25, 2026)
• GitHub Copilot plans (read Oct 25, 2026)
• GitHub Copilot CLI quickstart (read Oct 25, 2026)
• Gemini CLI README (read Oct 25, 2026)
• Antigravity CLI overview (read Oct 25, 2026)
• Amp pricing (read Oct 25, 2026)
• Qwen Code repository (read Oct 25, 2026)
• Kimi Code docs (read Oct 25, 2026)
• Devin CLI docs (read Oct 25, 2026)
• Pi coding agent README (read Oct 25, 2026)