There is no single best AI coding agent, because the categories do different jobs. Pick the category first: an editor if you read and edit code all day, a terminal agent if you hand over whole tasks, a cloud agent if you assign work and review it later. Then check what you already pay for. As of August 2026, Claude Pro at $20/mo includes Claude Code and ChatGPT Plus at $20/mo includes Codex, so a very large number of people already own a capable agent they have not opened.
- Pick the category first. Cross-category comparisons produce bad purchases.
- If you already pay for ChatGPT or Claude, you already own a terminal agent.
- Entry tiers cluster at $20 almost everywhere, so trying one is cheap.
- Every major agent now runs headless, which used to be the deciding feature.
- What happens at the limit differs enormously and is a daily experience.
- Most heavy users run two: an editor and a terminal agent, around $30 to $40 a month.
A pick per scenario
Start here. Find the row that describes your day, and the answer is the tool in the second column. The rest of this page explains why, and what each one costs when you push it.
Prices checked against vendor pricing pages in August 2026.
| If you | Use | Entry price |
|---|---|---|
| Read and edit code all day | Cursor | $20/mo |
| Want AI inside the editor you already use | GitHub Copilot | $10/mo |
| Already pay for ChatGPT | Codex | Included from Plus |
| Already pay for Claude | Claude Code | Included from Pro |
| Hand over whole tickets | Claude Code or Codex | $20/mo |
| Need agents inside CI | Any of them, with a key or token | Per token |
| Want to assign work and review it later | A cloud agent such as Devin | Free tier, $20/mo Pro |
| Want zero cost while you learn | Gemini CLI | Free, 1,000 requests/day |
| Want control and no vendor lock-in | Aider | Free plus token spend |
| Run several agents at once | Two subscriptions | About $40/mo |
The four questions
Answer these before you read a benchmark. They sort the field faster than any leaderboard, and unlike a leaderboard the answers stay true for years.
- Do you steer or delegate? Steering means you are reading the code as it changes and want the typing to be faster. Delegating means you describe an outcome and review a diff. Editors serve the first. Terminal agents serve the second.
- What do you already pay for? Claude Pro includes Claude Code. ChatGPT Plus includes Codex. Copilot Pro is $10. Checking your existing invoices is the highest-value ten minutes in this whole decision.
- Does anything need to run without you? Nightly jobs, PR triage, scheduled dependency bumps. This used to rule out editors entirely. As of August 2026 it does not, because Cursor and Copilot both ship command-line agents, but it still rules out the GUI-only way of working.
- What happens when you run out? Some tools stop until a window reopens. Some degrade to a cheaper model. Some keep going and bill you in arrears. This is the difference you will feel every week, and it gets almost no coverage.
Every price, checked in August 2026
Entry tiers have converged hard. Nearly everything serious starts at $20 a month, the free tiers are real rather than trials, and the top individual tier is $200 almost everywhere. That convergence is why price is a weak deciding factor and fit is a strong one.
Individual plans from each vendor’s own pricing page, read in August 2026. Vendors reprice without notice.
| Tool | Free tier | Entry | Mid | Top individual |
|---|---|---|---|---|
| Claude Code | No | $20 Pro | $100 Max 5x | $200 Max 20x |
| Codex | Yes, quick tasks | $20 Plus | No separate published rung | $200 Pro |
| Cursor | Yes, Hobby | $20 Pro | Pro+ price not published | Ultra price not published |
| GitHub Copilot | Yes, 2,000 completions/mo | $10 Pro | $39 Pro+ | $100 Max |
| Gemini CLI | Yes, 1,000 requests/day | Free | API key, per token | Vertex AI |
| Devin | Yes, light quota | $20 Pro | n/a | $200 Max |
| Aider | Software is free | Token spend only | Token spend only | Token spend only |
- Claude Pro is $17/mo billed annually and $20 month to month, and it is the cheapest legitimate route to Claude Code.
- ChatGPT Go is $8/mo and includes Codex for lightweight coding tasks only. Plus at $20 is the tier people actually mean.
- Copilot is free for verified students, and free for verified teachers and open-source maintainers. Nothing else in this table matches that.
- Team seats are not the same product. Claude Team is $25/seat month to month or $20 annually, Copilot Business is $19/seat, Cursor Teams is $40/user. Buy those for admin and billing, not for more capability.
What happens when you hit the limit
This is the question nobody answers and everybody experiences. Four different mechanisms are in play, and they produce four completely different Tuesday afternoons.
| Tool | Meter | What happens when it runs out |
|---|---|---|
| Claude Code | Usage window that resets every five hours, plus weekly limits on Max | You stop until the window reopens. Max carries two weekly caps, one across all models and one for Sonnet only. |
| Codex | Rolling five-hour window shared with cloud chats, plus a weekly cap | You stop, or you buy credits and continue at a published per-token rate. |
| Cursor | Included model usage per plan, then on-demand | It keeps working and bills you in arrears. Convenient, and the one that can surprise you. |
| GitHub Copilot | Unlimited completions on paid plans, plus monthly AI credits | Completions carry on. Premium model work stops until credits reset or you add more. |
| Gemini CLI | 60 requests/minute and 1,000 requests/day on a Google account | You wait, or attach an API key and pay per token. |
| Aider or any API key | None | Nothing stops. The bill grows. |
They all run headless now
For two years, "does it run in CI" was the cleanest line between an editor and an agent. That line is gone. As of August 2026 every tool below takes a prompt as an argument, does the work, prints, and exits.
claude -p "review the diff on this branch" --output-format json
codex exec "add a regression test for issue 412"
copilot -p "triage the failing job" -s --allow-tool 'shell(npm:*)'
agent -p "find and fix performance issues" --output-format text
gemini -p "explain the architecture of this codebase" --output-format json
aider --message "add type hints to utils.py" --yesWhat differs now is the guardrails around the run, not whether the run is possible. Those guardrails are worth more than the model in an unattended context, because an unattended agent with a bad plan and no ceiling is the expensive failure mode.
| Guardrail | Where it exists |
|---|---|
| Hard dollar ceiling per run | claude -p --max-budget-usd |
| Turn ceiling per run | claude -p --max-turns |
| Schema-validated JSON output | claude -p --json-schema, gemini --output-format json |
| OS-enforced sandbox | Codex, using Seatbelt on macOS and Landlock on Linux |
| Per-tool allowlist | copilot --allow-tool, claude --allowedTools |
| Auto-commit every edit | Aider, by default |
The two-tool answer
Most people who do this full time end up with an editor and a terminal agent, roughly $30 to $40 a month at entry tiers. That is not indecision. They serve different hours of the same day.
| Hour of the day | Tool |
|---|---|
| Reading unfamiliar code | Editor with AI |
| Making a change you can already see | Editor |
| A whole ticket you can describe in a paragraph | Terminal agent |
| Three tickets at once | Terminal agents, one git worktree each |
| Reviewing what came back | Editor |
| Anything scheduled or in CI | Terminal agent, headless |
How to decide in an afternoon
Check your invoices first
Open your billing pages for Anthropic, OpenAI, and GitHub. If you already pay for Claude Pro or ChatGPT Plus, install the agent it includes today and skip the rest of the shopping.
Install the free one anyway
Gemini CLI costs nothing and gives 1,000 requests a day on a personal Google account. It is a real terminal agent rather than a demo, so it doubles as a second opinion when your paid agent is stuck.
Give one agent a real ticket, not a toy
Pick something you were going to do anyway and that has a test. Toy prompts flatter every tool equally and tell you nothing about which one fits your repo.
Watch what happens at the limit
Run it hard enough in one sitting to hit a cap. How the tool behaves in that moment is the single most predictive thing you can learn in an afternoon.
Only then look at the editor
If you still find yourself typing most of the code, add Cursor or Copilot. If you find yourself describing outcomes instead, you already have what you need and the editor is optional.
Why a ranked list is the wrong shape
- The lead has inverted repeatedly. Any ranking is a snapshot of a race whose order changes with every model release, and there have been several this year.
- Benchmarks measure tasks that may not resemble yours. A two-point difference on a leaderboard is invisible next to whether the tool fits the way your repository is laid out.
- The category boundary moved. Editors grew CLIs, terminal agents grew GUIs and web surfaces, and vendors that shipped one product now ship four. A list ranked in 2025 is describing a market that no longer exists.
- Switching costs are low. Entry tiers are $20 and monthly. A month of the wrong choice costs less than a week spent researching the right one.
The durable advice is to choose for fit with how you work, because that changes slowly, rather than for a current benchmark, which does not.
Questions people ask
There is no single answer, because the categories serve different jobs. For editing code you are reading, Cursor. For handing over whole tickets, Claude Code or Codex. For assigning scoped work to run elsewhere, a cloud agent such as Devin. Pick the category first, then the tool.
Whichever comes with a subscription you already hold. As of August 2026, Claude Pro at $20/mo includes Claude Code and ChatGPT Plus at $20/mo includes Codex, so a great many people already own a capable terminal agent. Beyond that, GitHub Copilot Pro at $10/mo is the cheapest paid entry, and it is free for verified students.
Gemini CLI. It is Apache 2.0 licensed and gives 60 requests per minute and 1,000 requests per day on a personal Google account, which is a real working allowance rather than a trial. Copilot Free is the best free option if you want completions inside an editor instead.
Not necessarily, but most heavy users run two: an editor plus a terminal agent, around $30 to $40 a month at entry tiers. They cover different hours rather than competing for the same one.
All the major ones now. Claude Code takes -p, Codex has codex exec, GitHub Copilot CLI takes -p, Cursor CLI takes -p, Gemini CLI takes -p, and Aider takes --message. The differences are in the guardrails: budget ceilings, turn limits, and tool allowlists.
They address different work. Cursor makes the hours you spend inside code faster. Claude Code removes some of those hours. Comparing them head to head is a category error, which is why most people who use both keep both.
It depends on the tool, and this is the biggest experiential difference between them. Claude Code and Codex pause until a window reopens. Cursor keeps working and bills on-demand usage in arrears. Copilot keeps completions running and pauses premium model requests. An API key never stops at all.
Yes, repeatedly. The lead has already inverted several times and the category boundaries themselves moved this year. Choose for fit with your workflow, which is durable, rather than for a benchmark, which is not.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.