How to Pick an AI Code Assistant
Claude Code, GitHub Copilot, OpenAI Codex, or Gemini CLI? Real pricing, strengths, and limits as of August 2026, plus a decision table to pick yours.
Note
Before you start
- Any code editor or terminal you already use
- An account you can log into: GitHub (Copilot), Google (Gemini CLI), Anthropic or OpenAI (Claude Code / Codex)
- 10 minutes per tool to run the free trial paths in Step 5
Jump to section
- 1
Know the two camps
Inline IDE assistants (Copilot) complete code as you type; terminal agents (Claude Code, Codex, Gemini CLI) plan, edit, and run commands for you. Copilot now also offers its own agent and third-party agents.
- 2
Compare the four leaders
Claude Code (Anthropic), GitHub Copilot, OpenAI Codex, and Google's Gemini CLI - each with a free entry point and every price in the table verified against official pages today.
- 3
Match a tool to your workflow
Three realistic scenarios - IDE autocomplete, heavy agentic sessions, and budget-constrained or open-source-first setups - each with a concrete pick.
- 4
Try before you buy
All four have a $0 path: Copilot's Free tier, Gemini CLI with a Google account, limited Claude Code on the free plan, and Codex limited access. Trial each on a real task.
- 5
Commit for a month and check usage
Confirm billing terms match reality - completions vs credits vs token-based limits - and that the usage dashboard matches your actual pace before going annual.
Your AI code assistant is the single most used tool in your developer workflow - picking the right one is a budget decision, a model-quality decision, and a privacy decision rolled into one. This guide compares the four leading AI code assistants as of 13 August 2026 - Claude Code (Anthropic), GitHub Copilot, OpenAI Codex, and Google’s Gemini CLI - with every price and limit verified against official pricing pages today. You’ll learn the two camps they fall into, a side-by-side comparison table, three scenarios with a specific pick each, and the $0 trial path for every tool. For a head-to-head on the two most popular options, our comparison of Claude Code vs Copilot deep-dives further.
Before you start
- Know your workflow: do you want suggestions while you type (IDE assistant), an agent that plans and executes across files and the terminal, or both?
- Know your budget: every tool has a $0 entry point - the question is which paid tier you might grow into.
- Check the date: plan names, prices, and limits below are accurate as of 13 August 2026. All four vendors restructured pricing within the last year, so re-verify before paying.
- One rule that beats all four pricing pages: test the tool on a task you would actually do this week, not a toy example.
Step 1: Know the two camps
Every major AI code assistant falls into one of two camps, and the camps differ in how they help:
| Camp | How it works | Tools in this guide |
|---|---|---|
| Inline IDE assistant | Completes code as you type, chats about the open file, explains and refactors | GitHub Copilot |
| Terminal agent | Reads your repo, plans changes, edits files, runs commands and tests autonomously | Claude Code, OpenAI Codex, Gemini CLI |
The lines are blurring (official)
GitHub Copilot Pro now includes its own cloud agent - a terminal-like agent that works across your repos - plus access to third-party agents: Claude Code and Codex run inside Copilot on paid plans. And Claude Code and Codex both install inside VS Code too. The practical takeaway: start with the camp that fits your habit, not the UI.
Cursor and Windsurf, the well-known IDE-first assistants, deliberately sit outside this table - their pricing changes frequently, and we vouch only for sources we verified today.
Step 2: Compare the four leaders
The table below uses only official numbers, fetched today. Prices are per month; annual billing is cheaper where noted.
| Tool | Free option | Paid entry | Heavy use | Model / context |
|---|---|---|---|---|
| Claude Code (Anthropic) | Included in free plan (limited, 5-hour window) | Pro $17 (annual) or $20: full Claude Code + Cowork, 200k context | Max from $100 (5x/20x usage) | Sonnet, Opus, Haiku; 200K-token context |
| GitHub Copilot | $0: 2,000 completions/mo + 50 chats | Pro $10: unlimited completions, cloud agent, code review, third-party agents | Pro+ $39 / Max $100 | Model menu: Claude, GPT-5.x, Gemini, Grok, Kimi |
| OpenAI Codex | $0: limited Codex access | Plus $20: web, CLI, IDE, iOS, cloud integrations | Pro from $100 (5x or 20x limits) | GPT-5.6 family (Sol, Terra, Luna), GPT-5.4/5.5 |
| Gemini CLI (Google) | $0: 60 req/min, 1,000 req/day with Google account | same - API key route bills usage only | API key: pay per token | Gemini 3 models, 1M-token context |
How each plan actually bills (verified)
- Copilot completions vs AI credits: completions on the Free tier count against 2,000/month; Pro gives unlimited completions plus $15/month in AI credits (credits = $0.01 each) for agent runs, reviews, and premium-model chat.
- Codex credits: usage is measured in local messages, cloud chats, and code reviews per rolling 5-hour window, and extra usage draws from ChatGPT credits. API-key mode skips the cloud features and bills per token.
- Claude Code usage: included in all paid Anthropic plans, resetting on a rolling 5-hour window - you can burn it faster than you expect on big refactors, so check
/statusregularly. - Gemini CLI: the OAuth route with a personal Google account is free at 60 requests/min and 1,000 requests/day; the API-key route also starts free (1,000 requests/day) and scales by usage if you raise limits.
Step 3: Match a tool to your workflow
Three realistic scenarios, three specific picks:
Scenario A - “I live in my IDE and want smarter autocomplete.” Pick GitHub Copilot Free, then Pro ($10) when you hit the 2,000-completion ceiling - the cheapest unlimited option in this guide, and the only one with a third-party-agent switch (Claude Code or Codex inside your Copilot setup) if you grow into agentic work later.
Scenario B - “I want an agent that plans, edits, and tests across my whole repo.” Pick Claude Code if you want the strongest model menu and don’t mind the terminal (Pro at $17-20 includes it; our Claude Code setup guide gets you running in minutes). Pick Codex Plus ($20) if you already pay for ChatGPT - the cheapest full agent experience, and its cloud mode delegates heavy tasks while you sleep.
Scenario C - “Zero budget, open-source-first, or heavy Google user.” Pick Gemini CLI - Apache-2.0 licensed, runs entirely in your terminal, free at 1,000 requests/day with a Google account, and its 1M-token context window swallows entire codebases. It’s the only one of the four that is fully open source.
Not sure between Claude Code and Copilot?
Our side-by-side comparison covers model quality, permissions, price, and platform support in detail - the short version: Copilot wins on price and IDE reach, Claude Code wins on agentic capability and model flexibility.
Step 4: Try before you buy
Every tool has a verified $0 path. Run each candidate on the same real task (for example: “add an automated test for our auth module and wire it into CI”):
- Copilot: sign in with a GitHub account, pick Free - you get 2,000 completions and 50 chat requests monthly.
- Gemini CLI:
npx @google/gemini-cliand authenticate with your Google account - 1,000 requests/day, no card needed. - Claude Code:
npx @anthropic-ai/claude-code- free-plan access is limited by the 5-hour window, enough to evaluate; Pro ($17-20) unlocks full use. - Codex: use the Free ChatGPT plan’s limited Codex access, or
codexCLI on the $0 tier; upgrade to Go ($8) or Plus ($20) for real capacity.
Rate each trial on four questions: time-to-first-useful-result, how often it needed hand-holding, how close its edits came to your style, and how the limits felt.
Step 5: Commit for a month - then check usage
- Start monthly, not annual. All four vendors discount annual billing, but plan structures changed repeatedly in the last 18 months - month-to-month keeps you flexible.
- After a week, open the usage dashboard (
/statusin Claude Code and Codex CLI; Copilot’s usage panel; Gemini CLI’s quota view) and compare your actual pace against the limits. - Watch the three billing traps: credits that aren’t completions (Copilot), 5-hour rolling windows that reset mid-task (Claude Code, Codex), and “5x/20x more usage” that multiplies limits, not model quality.
- Re-evaluate quarterly. The July 2026 Codex Pro restructure and Copilot’s credit system both landed within the last year - pricing pages are moving targets, and this table has a shelf life.
How to verify it worked
- You can explain the difference between an inline IDE assistant and a terminal agent
- You named a price for the free tier of all four tools ($0, $0, $0, $0)
- Your scenario has a named pick with a one-sentence justification
- You ran at least one trial on a real task and scored it on the four questions
- You checked your chosen plan’s usage dashboard and know your pacing
- You know which of the four camps the other options (Cursor, Windsurf) belong to
Troubleshooting
I can’t tell “AI credits” from “included usage” on Copilot
Completions and chat on the Free tier count against your 2,000-completion monthly block and stop there. Pro includes unlimited completions plus $15/month of AI credits (1 credit = $0.01) that agent runs, code reviews, and premium-model requests draw from - think of credits as the paid battery and completions as free solar. The usage panel shows both meters separately; when the credit meter empties you either wait for the monthly reset or buy more.
Gemini CLI says “rate limit exceeded” on the free tier
The OAuth route caps at 60 requests per minute and 1,000 per day with a personal Google account. Options: wait for the window to reset, switch to the Gemini API-key route (also free at 1,000 requests/day for Gemini 3, paying only when you raise limits), or use a longer single session instead of many short ones - each turn in a session consumes one request.
Codex says “you’ve hit your usage limit” mid-task
On Plus, usage is measured in local messages and cloud chats per rolling 5-hour window - a long multi-file task can consume many times the allowance of a short edit (model choice, context, and tool use all count). Continue with ChatGPT credits if you bought them, upgrade tier, or let the window roll over. Free and Go plans have the tightest caps, so Plus remains the realistic floor for daily Codex work.
Claude Code burns through my 5-hour window on one refactor
Expected behavior: the window is usage-based, resets on a rolling 5-hour clock, and big agentic jobs dominate. Run claude /status to watch consumption, split huge tasks into smaller requests, and consider --permission-mode to cut autopilot waste. Only the $100 Max tier’s 5x-20x usage changes the math materially.
A tool’s website shows different prices than this table
Normal - and expected. All four vendors changed plans within the last year (OpenAI restructured Codex pricing and credits in April 2026, Copilot introduced its credit system in early 2026, Anthropic reworked Pro/Max tiers). This table was fetched 13 August 2026; when a page drifts, trust the official page over this guide and re-run the trial in Step 4.
You did it
- You know the two camps and where all four tools sit
- You verified the $0 entry points and paid tiers yourself on the official pages
- You picked a tool for your scenario - and you can justify it in one sentence
- You ran real-task trials before paying for anything
- You understand the billing model of your pick (completions, credits, windows, or tokens)
- Pricing is a moving target - you know to re-check before you commit annually
Official sources
- Anthropic pricing (Pro/Max/Team/Enterprise, Claude Code included, API rates): anthropic.com/pricing
- GitHub Copilot plans (Free/Pro/Pro+/Max, completions, credits, third-party agents): github.com/features/copilot/plans
- OpenAI Codex pricing (Free/Go/Plus/Pro/Business, usage limits, credits, API key): developers.openai.com/codex/pricing
- Gemini CLI repository (Apache-2.0, free tier limits, install, API key): github.com/google-gemini/gemini-cli
All fetched 13 August 2026. Plan names, prices, and limits change frequently - verify before purchase.
Next: installed Claude Code and want the details? Set it up properly, or read the Claude Code vs Copilot comparison before committing. And once your assistant feels at home, wire it to the wider MCP ecosystem.
Questions, answered first
Which AI code assistant should I pick?
Match the tool to the workflow: want autocomplete in your IDE at $0 - GitHub Copilot Free. Want a terminal agent with the strongest models - Claude Code (Pro at $17-20/month includes it, Max from $100 for heavy use). Already paying for ChatGPT - Codex is bundled into Plus at $20/month. Want fully free open-source terminal agenting with a Google account - Gemini CLI.
Are these really free to try?
Yes, all four have verified $0 entry points: Copilot Free (2,000 completions/month), Gemini CLI (1,000 requests/day with a personal Google account), limited Claude Code access on Anthropic's free plan, and Codex's limited free access. Each is enough for a weekend trial on a real project - which is exactly the test this guide asks you to run before paying.
What does '5x or 20x more usage' on Codex Pro and Claude Max actually mean?
It multiplies the included usage of the base paid plan, not the intelligence of the model. Codex Pro 5x ($100/month) gives 5x the Plus usage limits; the $200 tier gives 20x. Claude Max scales usage in a similar spirit - about 5x on the $100 tier and 20x on the $200 tier relative to Pro. If you rarely hit limits on the base plan, the expensive tier buys nothing.
Do I need to worry about my code or prompts leaving my machine?
With all four, prompts and code you send to the service go to that company's servers for processing - that's the tradeoff of cloud coding. Your streaming/telemetry settings are configurable per tool (for example, Claude Code's permission modes and Gemini CLI running fully in your terminal), and the local-first alternative - models on your own hardware - is covered in our local LLM guide.
What about Cursor, Windsurf, or other assistants?
This guide covers the four assistants with official, verifiable pricing pages: Claude Code, GitHub Copilot, OpenAI Codex, and Gemini CLI. Cursor and Windsurf are capable IDE-first assistants - the same 'try on a real task before paying' rule applies - but their pricing changes frequently and we deliberately vouch only for sources we verified today.
You did it
- You can name the two camps and which camp each of the four tools belongs to
- You compared starting prices: $0 everywhere, $10 Copilot Pro, $17-20 Claude Pro, $20 Codex Plus
- You picked a tool for your scenario and can justify it in one sentence
- You ran at least one free trial on a real task, not a toy example
- You know how your chosen plan bills: completions (Copilot Free), credits, or tokens
- You know where your usage data lives - GitHub, Anthropic, OpenAI, or Google