AI Coding Assistants Compared¶

In one line: Pick a coding assistant by where you work (terminal, IDE, or delegated cloud tasks), which models you want, and how billing works, then judge it on a real week of your own tasks rather than on benchmarks. Last reviewed: 2026-10-21

The short version¶

• Coding assistants now come in three modes: autocomplete (suggests the next lines), chat and agent in your editor or terminal (reads your repo, edits files, runs commands), and cloud or background agents (you hand off a task and get back a pull request).
• As of 2026-10 the main options are Claude Code, GitHub Copilot, Cursor, OpenAI Codex, Google's Gemini CLI and related tools, and open-source agents such as OpenCode and Aider that work with any model.
• Prices are mostly a flat monthly plan with usage limits or credits on top. Heavy agent use can hit limits fast. Read the usage rules, not just the headline price.
• Pick Claude Code if you live in the terminal and want a strong autonomous agent; Copilot if your team is on GitHub and wants the cheapest broad option with admin controls; Cursor if you want an AI-first editor (note its 2026 ownership change and Grok-first plans); Codex if you already pay for ChatGPT; Gemini CLI for a free, open-source starting point; OpenCode or Aider if you want to bring your own model or run locally.
• Whatever you choose: review every diff, keep secrets out of prompts, and run the tests. The assistant writes code; you own it.

What these tools actually do¶

flowchart LR
    A[Autocomplete: next lines as you type] --> B[Chat: explain, refactor, write a function]
    B --> C[Agent in editor or terminal: reads repo, edits many files, runs tests]
    C --> D[Background / cloud agent: works alone, opens a pull request]

Moving right, the tool does more on its own, saves more time on well-specified tasks, and needs more careful review. The research behind this progress: Codex (2021) for code models, SWE-bench for measuring agents on real GitHub issues, and METR time horizons for how long a task agents can finish. Benchmarks and their limits: maths and code benchmarks.

Comparison table (as of 2026-10)¶

Notes on the table:

• Cursor ownership changed in 2026. Press reports say SpaceX's AI unit (which absorbed xAI) completed an acquisition of Cursor's parent Anysphere in August 2026, and Cursor's current plans centre on Grok models with other models billed at API rates. If you chose Cursor for access to a particular non-Grok model, re-check what your plan includes.
• Google notes on the Gemini CLI repository that the CLI is not optimised for its newest Gemini 4 model and points users to a newer "Antigravity" CLI. Expect Google's lineup here to keep shifting.
• Copilot's free and paid quotas, and every plan above, change often. Check the source links before buying for a team.

How to choose¶

Pick Claude Code if you are comfortable in a terminal, want an agent that can plan and carry out large changes across a repo, and want to extend it with MCP servers and custom commands. It is also the most scriptable: the same harness is available as the Claude Agent SDK. See build a Claude agent with MCP.

Pick GitHub Copilot if your organisation already uses GitHub and wants one approved tool for everyone, across many IDEs, with central policy and the lowest entry price. Copilot code review and its cloud agent work directly on pull requests.

Pick Cursor if you want the editor itself to be built around AI (inline edits, multi-file changes, cloud agents) and you are happy with Grok models or paying API rates for others.

Pick OpenAI Codex if you or your company already pay for ChatGPT and want to try an agent without another subscription.

Pick Gemini CLI if you want a free, open-source terminal agent to learn on, or you are on Google Cloud.

Pick OpenCode or Aider if you need to choose your own model, run against a local model for privacy, or avoid lock-in. Expect more setup and weaker results with small local models.

Many developers use two: an IDE assistant for everyday edits and a terminal or cloud agent for larger tasks.

How to evaluate one in a week¶

Benchmarks tell you little about your codebase. Run your own trial:

1. Pick 10 real tasks from your backlog: two bug fixes, two small features, a refactor, tests for untested code, a dependency upgrade, a documentation task, an unfamiliar-codebase question, and one hard task.
2. Run each tool on the same tasks. Give the same instructions.
3. Score: did it finish? How many corrections did you make? Did the tests pass? Any security or quality issues in review? Time spent, including review.
4. Check cost: how far into the plan's limits did the week take you?
5. Decide on finished-and-correct tasks per hour of your time, not on how impressive the demo felt.

Safe use¶

• Review every diff. Treat agent output like a pull request from a fast, confident junior colleague.
• Keep secrets out. Do not paste credentials into chats; check what files the agent can read; use .gitignore-style ignore files the tool supports.
• Limit permissions. Agents that run shell commands should ask before destructive commands. Run them in containers or dev environments, not on production machines.
• Watch for prompt injection. A README, issue or web page the agent reads can contain instructions. See agent security.
• Check licences and policy. Your employer may have rules on which tools and models are allowed. See acceptable use policy and coding assistants in practice.
• Keep learning the fundamentals. If you are early in your career, make sure you can explain every line you ship. See AI engineer roadmap.

Common mistakes¶

• Judging by the demo. Run your own tasks.
• Accepting large diffs unread. Bugs and security issues hide in volume.
• Ignoring usage limits. Agent-heavy weeks can exhaust a plan. Know what happens at the limit (throttling, pay-as-you-go, or a stop).
• One giant prompt. Agents do better with a clear goal, constraints, and how to verify (for example, "make the tests in tests/api pass").
• No project instructions file. Most tools read a repository instructions file (for example CLAUDE.md or AGENTS.md). A short one with build commands and conventions improves results a lot.

Next steps¶

• Coding assistants in practice and the engineers' playbook
• Agent architectures to understand what these tools do inside
• Chat assistants and local LLM runners
• Which model for which job

Sources¶

• Anthropic, Claude plans and pricing (Pro, Max, Team, Enterprise; Claude Code included; as of 2026-10): https://claude.com/pricing
• GitHub, Copilot plans (Free limits, Pro, Pro+, Max, AI Credits, Business/Enterprise; as of 2026-10): https://github.com/features/copilot/plans
• Cursor, Pricing (Free, Plus, Premium, Super, Ultra, Teams; as of 2026-10): https://cursor.com/pricing
• tbreak, report on Grok 4.6 in Cursor and the SpaceXAI acquisition of Anysphere (closed 14 August 2026): https://tbreak.com/grok-4-6-cursor-spacexai-acquisition/
• Google, Gemini CLI repository (Apache 2.0, install, Antigravity CLI note; as of 2026-10): https://github.com/google-gemini/gemini-cli
• OpenAI, Codex CLI repository: https://github.com/openai/codex
• OpenCode: https://github.com/sst/opencode ; Aider: https://github.com/Aider-AI/aider