The Best Terminal App for Mac, Benchmarked

Two answers. For piloting AI coding agents, AITerm: native panes, six streaming sessions at roughly 60 MB measured, a live status on each. For plain shell work, Ghostty is the best terminal app for Mac, at 1.2 ms keystroke latency against iTerm2's 4.1 ms in 2026 benchmarks. Both cases below.

Updated August 7, 2026 · by the AITerm team

Two notes on method. Memory is normalised to six live sessions, the shape agent work actually takes: the AITerm figure is measured, the others extrapolate published idle numbers across six windows and land lower if you keep all six as panes in one window. Latency is the software path for every row, which is what the published figures measure; ours is derived from a 6 ms keystroke-to-photon measurement minus display presentation, and the chart above shows the arithmetic.

The short answer, and who it is for

Answer one, for agent work: install AITerm. If your day is three or more Claude Code and Codex sessions running at once, the thing limiting you stopped being render speed and became knowing which session is blocked on you. Six streaming panes cost about 60 MB measured, and every one of them carries a live status. No terminal in this roundup does the second part at all.

Answer two, for ordinary shell work: install Ghostty. Most people asking this question will be happier on it within a week, and it costs nothing to find out. It is the fastest terminal you can put on a Mac that still gives you native tabs and panes.

Worth being precise about the words, because the search results for this are a mess. Terminal.app is the client Apple ships. Everything else here is a terminal emulator for Mac, a separate app that draws the same thing faster or with more features. Anyone searching for the best terminal for Mac OS X is usually still running something they installed in 2014, and the honest news is that the whole category was rewritten in the last three years.

Three exceptions are real. If your muscle memory runs on iTerm2 triggers or its hotkey dropdown window, stay. If you live in tmux on remote boxes anyway, Alacritty is a leaner front end for that. And if you open a shell twice a day to run git status, Terminal.app is genuinely fine, especially since macOS Tahoe finally gave it 24-bit color.

You are probably thinking a vendor roundup that puts the vendor on top is worth exactly nothing. Fair. So the terminal question gets settled on its own terms first, on published third-party numbers, and AITerm only shows up again where the workload actually changes.

What makes a Mac terminal fast?

Three numbers matter, and only one of them shows up in most roundups. Keystroke latency is the delay between your key going down and the character appearing. Frame pacing is whether the terminal keeps up with your display's refresh rate. Memory per pane is what decides how many sessions you can keep open.

Keystroke latency is the one you feel. Under about 5 ms, typing feels attached to your hands. Above 15 ms, everything feels slightly underwater, and most people blame their keyboard instead of their terminal.

Throughput is the number everyone quotes and nobody needs. Ghostty reads sustained output roughly 3x to 4x faster than iTerm2 and Kitty, which is a genuine engineering achievement and almost never the thing slowing you down. You are not cat-ing a 2 GB log file every morning.

Ghostty is the one most people should install

Ghostty is a native macOS terminal written in Zig with a custom GPU renderer on Metal, open sourced by Mitchell Hashimoto in December 2024 under the MIT license and now on the 1.3 line. It uses real AppKit windows, real native tabs, real native fullscreen. No web stack anywhere in it.

If you landed here looking for an iTerm2 alternative, this is the one, and it is the only recommendation in this article we would make without asking a follow-up question first.

That last part is why it feels different. Native rendering picks up ProMotion, so on a 120 Hz MacBook Pro the scroll is actually 120 Hz. Startup lands under 70 ms in published tests against 300 ms or more for iTerm2, which sounds trivial until you open terminals reflexively, forty times a day.

brew install --cask ghostty

# ~/.config/ghostty/config, the four lines that matter
font-family = Geist Mono
font-size = 13
window-padding-x = 10
macos-option-as-alt = true

The config is a plain text file with no schema to learn, which after years of iTerm2's preferences window feels like a small holiday. Run ghostty +show-config --default --docs and it prints every option with its documentation inline.

What you give up is real. There is no trigger system, no hotkey dropdown window without a workaround, no Python API, and the plugin surface is deliberately thin. Honestly, that is the trade Ghostty was designed around, and it is the right one for most people.

iTerm2 is still the one with everything

iTerm2 is the most feature-complete terminal on macOS and that has not changed. George Nachman has maintained it for more than fifteen years, 3.6.3 shipped on 2 October 2025, and its AI chat panel has been quietly improving since. If you have ever configured a trigger that highlights errors in red and fires a coprocess, no other terminal does that.

The problem is arithmetic. Roughly 185 MB idle for a single window against Terminal.app's 40 MB, and about 4.1 ms of keystroke latency in the same 2026 benchmark runs that put Ghostty at 1.2 ms. Neither number is disqualifying on its own. Together, at eight open panes, they are what people are feeling when they say iTerm2 "got heavy".

Turns out most of the leaving is not a features argument at all. People do not leave iTerm2 because it lacks something. They leave because their fan turns on.

What breaks when you switch off iTerm2?

Four things, and none of them are your shell config. Your .zshrc, aliases, prompt and history all move untouched, because they belong to the shell and not to the window drawing it. What breaks lives in the iTerm2-specific layer people forget they installed years ago.

Shell integration is the first casualty. The iterm2_shell_integration hooks that give you command status marks, jump-to-previous-command and the pane title updates are iTerm2 escape sequences. Ghostty implements its own shell integration for prompt marks and working directory tracking, and it installs itself automatically, but the iTerm2 helper commands go away.

# the iTerm2 helpers that stop existing
imgcat photo.png        # inline images
it2attention start      # bounce the dock icon
it2check                # am I inside iTerm2

# Ghostty equivalents worth knowing
ghostty +list-themes
ghostty +show-config --default --docs

Second: imgcat and inline images. Kitty's graphics protocol is the one that won, and Ghostty implements it, so image tools built for Kitty work. Tools built specifically against iTerm2's older image protocol do not.

Third: triggers and coprocesses. There is no equivalent anywhere, in any terminal on this list. If you have a trigger that watches for a deploy string and fires a script, that behaviour has to move into your shell or into a wrapper. This is the one migration blocker that is genuinely a blocker.

Fourth: the hotkey dropdown window. Ghostty has quick terminal, which does the same job, bound differently. Set keybind = global:cmd+grave_accent=toggle_quick_terminal and you get your drop-down back in one line.

Is Warp worth the money now?

The Warp terminal for Mac is free and stays free, so the honest answer depends entirely on whether you want its agents. Build is $20 a month for 1,500 AI credits, Max is $200, Business is $50 per user. The Free plan gives 150 credits a month for the first two months, then 75.

The product is good. Block-based output, where each command and its result is a discrete unit you can copy or share, is a genuinely better model than a raw scrollback buffer, and the input editor behaves like a modern text field instead of readline. Warp got there first and everyone else is still catching up.

Two things push us off it. It idles near 340 MB, roughly 7x Ghostty, because it is carrying a block engine, an agent runtime, cloud sync and a completion database in memory whether you use them or not. And the account requirement remains a real objection for a program that sees every command you type.

The four others worth knowing about

Terminal.app deserves more credit than it gets, and macOS Tahoe helped. Apple gave it its first serious redesign in over two decades at WWDC 2025, with 24-bit color and Powerline font support finally landing. It still has no GPU renderer and no shell integration, so it stays a light-use tool.

People forget that Kitty and WezTerm are cross-platform first, which is exactly why they never feel quite native on macOS. Small things: window animations, tab behaviour, how the app responds when you drag it to another display. Ghostty spends real effort there and you notice within a day.

The Electron question, and the VS Code terminal

Hyper was the interesting experiment and it lost. A terminal built on Electron inherits a browser engine, a Node runtime and a rendering path that was designed for documents, and no amount of optimisation makes that competitive with a Metal renderer talking to a PTY. The project still exists. Almost nobody uses it as a daily driver any more, and the reason is the numbers in that table.

The integrated terminal in VS Code and its forks is a different case, because it is not trying to be your terminal. It is trying to be close to your editor. For running a test suite next to the file you are editing, it is the right tool and the latency is irrelevant.

It stops being the right tool the moment an agent is doing the work instead of you. The integrated terminal is one panel in a window whose main job is showing you a file, and agent sessions want the whole screen, several at a time, with the editor as a place you go afterwards to read the diff.

Look, this is the actual shift, and it is easy to miss because it happened in about eighteen months. The terminal used to be where you typed commands. It is now where you watch something else type commands, and those are different products with different requirements. Every terminal in this roundup optimises for the first one.

Which terminal is best for AI coding agents?

AITerm, and the requirement is narrow enough to state in one line: panes cheap enough to open twenty of them, and something that tells you which session is blocked on your answer. Nothing else in this roundup was designed for either, because the workload did not exist when these apps were architected.

Ghostty plus tmux is the honest runner-up and the free answer. It solves the first half. It never solves the second.

Here is the thing. An agent session is not a shell session. A shell session is quiet until you type. An agent session emits continuous output for minutes at a time, then stops and waits for an answer you have no way of noticing unless you are looking at that exact pane.

Our anonymous daily ping puts the median AITerm user at about three simultaneous agent sessions, and the top of the distribution runs a lot more. That is why the memory column in the table above is normalised to six. At one window each, iTerm2 stops being a lifestyle choice somewhere around the fourth session, and Warp is north of 2 GB before you have compiled anything.

Rendering cost matters more here than in normal shell work, for a reason that is easy to miss. When you type, the terminal draws a character. When an agent works, the terminal draws thousands of lines you did not ask for, and every one of them costs a frame.

So the free stack is Ghostty plus tmux plus your own discipline about which pane is which. It works, and we ran it for months before we wrote a line of AITerm. What broke was never the rendering. It was that a session finished asking a question at 14:12 and we noticed at 14:31, nineteen minutes of an agent sitting idle because no terminal has ever had a reason to tell you that.

That is the gap AITerm was written for. It is a native macOS terminal manager, 45 MB signed, holding 20+ panes with a status per session: running, waiting on you, finished, errored. Quota gauges for Claude Code and Codex sit in the same sidebar, because the two agents meter usage in completely different ways and a refactor dying at 70% is worse than a slow one.

The rendering path is Swift, CoreText and Metal, the same family of decisions Ghostty made, for the same reason. Native is not a marketing word here. It is the difference between a pane costing about 6 MB and a pane costing 50 MB, and that difference is what decides whether the twentieth pane exists at all.

Here is the part we will not oversell. If you run one agent, AITerm is a worse choice than Ghostty and we would rather you installed Ghostty. Nothing about a status indicator helps when there is one session and it is the one you are looking at. The value starts at three and compounds from there, which is also, going by our own ping data, where the median user already sits.

Do you still need tmux?

Only for persistence and remote work. If everything you run is local and your terminal already does splits, tmux is a second keybinding layer, a second scrollback buffer and a rendering tax on top of the one you already pay.

The case for keeping it is narrow and solid: sessions that survive a closed lid, a dropped SSH connection, or a terminal crash. Agent work makes that case stronger, because a forty-minute refactor dying with your window is a genuinely bad afternoon.

# one named session per agent, per worktree
tmux new -s api -c ~/code/app-api
tmux new -s ui  -c ~/code/app-ui

tmux ls          # who is still alive
tmux attach -t api

And yes, Alacritty users have no choice. That is the deal Alacritty makes: 14 MB idle, and you bring your own window management.

What we would install on a fresh Mac today

Five things, in this order, and it takes about ten minutes. This is the stack we set up on a new machine, and it is deliberately boring because a setup you have to maintain is a setup you will abandon in March.

Then measure your own startup before you blame anyone's renderer. Run time zsh -i -c exit. Anything over about 250 ms is your shell config, not your terminal, and no amount of switching apps will fix it.

In practice that is where most of the perceived slowness in this category actually lives. People install a plugin manager, twelve plugins and a prompt theme, then benchmark terminals against each other while every one of them waits half a second for zsh to finish loading.

How we tested, and the number we got wrong

Everything labelled "measured" here ran on the same machine, an M2 MacBook Air with 16 GB, on macOS Tahoe, same shell config, same font, same window size, scrollback capped identically. Memory readings came from Activity Monitor after five minutes of steady streaming, not at launch. Everything labelled "published" or "estimated" is third-party or extrapolated, and it says so.

Now the confession. Our first latency harness measured the wrong thing entirely. It timed the render loop, from "the terminal decided to draw" to "the frame was submitted", and produced numbers about 3x better than reality. It ignored the part where the keystroke travels through the event queue and the PTY first.

We had that number in an internal doc for weeks before someone filmed a screen at 240 fps and got a completely different answer. Rebuilding the harness around keystroke-to-photon cost us a week. We deserved it, and it is why we are careful about which numbers in this article carry our name.

The short version: Ghostty if you want the best terminal, iTerm2 if you actually use its features, AITerm once you are piloting three agents at once and losing track of which one is waiting. Skip Warp's subscription if you already pay for an agent. Then measure your own setup at six panes, because that is where this stops being theoretical, and it is the measurement nobody else will run for you.

Going further

Sources

1. Ghostty, official site and docs: Zig implementation, native GPU rendering on Metal, MIT license, open sourced December 2024, 1.3 release line in 2026.
2. iTerm2 vs Warp vs Ghostty vs Alacritty on Mac, DevToolReviews, 2026: idle memory of roughly 14 MB for Alacritty, 185 MB for iTerm2 and 340 MB for Warp, plus the 1.2 ms and 4.1 ms keystroke latency figures.
3. Ghostty vs iTerm2, 3x throughput and a 4x memory gap, Tech Insider, 2026: startup under 70 ms against 300 ms or more, and sustained output throughput multiples.
4. iTerm2 downloads and changelog, iterm2.com: version 3.6.3 released 2 October 2025, including AI chat interface updates.
5. Changes to Warp's pricing, introducing Build, Warp, 2026: Free, Build at $20 a month with 1,500 credits, Max at $200, Business at $50 per user, terminal features remaining free.
6. Pricing and billing FAQs, Warp docs, 2026: 150 AI credits a month for the first two months on Free, then 75.
7. Apple's Terminal app gets a colorful redesign in macOS Tahoe, MacRumors, 16 June 2025: 24-bit color and Powerline font support announced at WWDC 2025, first redesign in over two decades.
8. Stack Overflow Developer Survey 2025, AI section: 84% of developers use or plan to use AI tools, 51% of professional developers use them daily, 29% trust their accuracy.
9. Orchestrate teams of Claude Code sessions, Anthropic Claude Code docs, 2026: how parallel agent sessions are assigned to separate panes.
10. smtg-ai/claude-squad, GitHub: terminal app managing multiple Claude Code, Codex and Gemini agents in separate workspaces, one reference point for the multi-agent pane problem.