AI coding agents compared: what each one is actually good at

The short answer

Every serious AI coding agent now costs about $20 a month, runs roughly the same frontier models, and plugs into roughly the same editors. That convergence is real, it happened fast, and it means most comparisons of these tools are comparing the wrong things.

This guide is built on what we could verify today — 14 August 2026 — from vendors' own pricing pages, billing documentation and GitHub. It is deliberately not a tour of ten products. It answers three questions: what your money actually buys, what genuinely still differs between these tools, and which one fits the job you have.

Everyone charges the same now. That is the actual news.

Put the four biggest vendors' price ladders side by side and they are nearly identical:

• Anthropic: Pro $20/month ($17 billed annually), Max from $100/month (claude.com/pricing)
• OpenAI Codex: Free, Go $8, Plus $20, Pro from $100 for 5x limits or $200 for 20x (learn.chatgpt.com/docs/pricing)
• Cursor: Hobby free, Pro $20, Pro+ at 3x Pro's agent limits, Ultra at 20x (cursor.com/pricing)
• GitHub Copilot: Free, Pro $10, Pro+ $39, Max $100 (GitHub docs)

The same rungs appear everywhere: a free trial, a $20 entry tier, roughly $100 for five times as much, roughly $200 for twenty times as much. Cursor no longer even describes its upper tiers in terms of features — Pro+ is "3x Pro limits on Agent" and Ultra is "20x". You are not buying capabilities. You are buying a token budget and a place to spend it.

Which means price comparison between vendors is close to useless, and the question worth asking instead is what a given budget actually gets you.

What $100 actually buys

GitHub is the only vendor that publishes this plainly, and the numbers are worth reading carefully. One AI credit is $0.01, credits are consumed at token rates rather than per request, and the allowances are (GitHub billing docs):

Every tier hands back more metered usage than it charges, and the gap widens as you go up: 1.5x at Pro, about 1.8x at Pro+, exactly 2x at Max. On top of that, code completions and next-edit suggestions are "not billed in AI credits and remain unlimited for all paid plans" — the highest-frequency feature is simply not metered.

Now set that against how much a working developer actually consumes. Anthropic's own cost documentation puts real deployments at around $13 per developer per active day, and $150–250 per developer per month, with 90% of users staying under $30 per active day (Claude Code docs). That is the most useful number in this entire article, and it comes from a vendor describing its own customers.

Read those two facts together and the conclusion is uncomfortable for the marketing: a $20 plan is a trial, not a budget. Professional daily use lands in a band that a $20 tier cannot cover, at any vendor. If you intend to run agents all day, Copilot Max at $100 for $200 of usage is the most subsidised per-seat token budget we could verify anywhere.

One exception is worth knowing about. Augment Code's Business plan is $100/month flat with $100 of usage, up to 50 seats, no per-seat charge (augmentcode.com/pricing). Everyone else bills per seat. That inverts the maths for a particular shape of team: if most of your seats are occasional users and two or three people do the heavy lifting, pooling is far cheaper than buying everyone a Max seat. If your whole team runs agents all day, you will exhaust the pool early and buy top-ups at pay-as-you-go, and the per-seat subsidy elsewhere wins. The rough crossover is a team average of about $100 per developer per month — below Anthropic's observed band, which tells you pooling suits mixed teams rather than uniformly heavy ones. (The units are not perfectly comparable: Augment's allowance covers its context engine and compute as well as model calls.)

Augment also carries SOC 2 Type II, ISO 42001 and CMEK, which is the practical reason it shows up in regulated shortlists.

The model is not what you are choosing any more

This is the part that has changed most, and it should change how you shortlist.

GitHub's own Copilot rate card lists Claude Opus 5 and Claude Sonnet 5 alongside its GPT-5 family (models and pricing). You can run Anthropic's frontier models inside Microsoft's product. Cursor sells "access to frontier models" as a Pro feature. Cline and Aider have always been bring-your-own-key. Whatever model you think you are choosing a tool for, you can probably reach it from several of these tools.

The editor lock-in went the same way. The Agent Client Protocol — an open standard for connecting agents to editors, in the way LSP connected language servers — now lists 50+ agents and 12 editors, including Claude Agent, Cline, Cursor, Devin, GitHub Copilot, OpenHands, Codex CLI and Gemini CLI on one side, and Zed, JetBrains IDEs, VS Code, Neovim, Emacs and Obsidian on the other (zed.dev/acp).

So "which editor" and "which agent" are no longer the same decision. If you have been avoiding a tool because you did not want to leave your editor, check whether that constraint still exists before you rule it out.

The clearest sign of where this leads is OpenHands. Its repository headline is no longer an agent at all — it is Agent Canvas, described as "the self-hosted developer control center for coding agents and automations," designed to run "OpenHands, Claude Code, Codex, Gemini, or any ACP-compatible agent" across local, remote and cloud backends (github.com/OpenHands/OpenHands). The agent is becoming the commodity. The harness is becoming the product.

What genuinely still differs

Strip out model and editor and three things are left, and these are what you should actually compare.

Where it runs. Claude Code and Aider run in your terminal, on your machine. Copilot's cloud agent and Devin run work on vendor infrastructure rather than your laptop. OpenHands runs wherever you put it, including entirely inside your own infrastructure — which is the whole answer for teams whose code cannot go to a third party.

Who holds the token bill. A subscription means the vendor absorbs variance and caps your exposure. Bring-your-own-key means you pay raw API rates with no markup and no cap. At light use, BYOK is cheaper. At Anthropic's observed $150–250 per developer per month, a subsidised plan like Copilot Max is usually cheaper than paying list API rates yourself — the subsidy is real money.

How long a leash it has. This is the genuine capability spread, and it is what people mean when they say one agent is "better". Copilot's completions sit at one end: constant, low-stakes, unmetered. Claude Code and Devin sit at the other, running for long stretches and touching many files. A tool that is excellent at one end is not automatically good at the other, and most teams need both.

The open-source options, and whether anyone is still shipping

Star counts are the most quoted and least useful number in this category. Here is what we pulled from the GitHub API today:

MetaGPT is the second most-starred project here and has not taken a commit since January. Aider is the one that needs the most careful reading: it is a genuinely well-designed tool, its git-native commit workflow is still the cleanest of any CLI agent, and it was actively committed to as recently as May — but its last tagged release, v0.86.0, was 9 August 2025, over a year ago, against 1,797 open issues. If you are standardising a team on Aider, confirm its current model support yourself first.

Cline and OpenHands both cut releases yesterday (v4.1.9 and v1.13.0 respectively). If you want an open-source tool that will still support whatever model ships next month, those are the two with the evidence behind them.

SWE-agent is worth naming for what it is: a research and benchmark harness with a tight, well-maintained codebase, not a daily driver.

Three you can stop evaluating

These still appear on comparison lists, including some published this year:

• Kite. Discontinued. kite.com has served a farewell notice since 2021. (It is still in our own index as we publish this — we are fixing that.)
• Sourcegraph Cody. Cody Free and Cody Pro were discontinued on 23 July 2025. Cody continues for Sourcegraph Enterprise customers; Sourcegraph's product for individual developers is now Amp.
• Amazon Q Developer's IDE plugins. AWS states it "will discontinue support for Amazon Q Developer IDE plugins" on 30 April 2027 and points developers to Kiro (aws.amazon.com/q/developer). Q Developer elsewhere in AWS is unaffected. If you are an AWS shop starting fresh, evaluate Kiro rather than adopting the plugins now.

How to actually decide

If you take one thing from this: shortlist on where the agent runs and who pays for the tokens, not on which model it uses or which editor it lives in. Those last two have largely stopped being differentiators.

Then do the arithmetic honestly. Estimate your real monthly model spend — Anthropic's $150–250 per developer is the best public anchor — and compare it against what each plan gives back. At that level of use, the tools that subsidise metered usage cost less than the tools that do not, and the difference is larger than any feature gap between them.

And give whatever you pick a real task from your own codebase for a week. Every claim in this article is checkable; how a tool feels on your code is not.

We track this category as it moves in our coding agents index, and our longer tool-by-tool roundup is at best AI coding agents.

What we could not verify

Cursor does not render prices for Pro+, Ultra or Teams Premium in its page source, so we have quoted only the usage multiples it states rather than figures we could not confirm. We did not independently test any tool's coding quality, and we deliberately quote no benchmark scores — the public leaderboards are contested enough that a single number would mislead more than it informs. Claude Code's plan requirements are taken from Anthropic's setup documentation, which states it "requires a Pro, Max, Team, Enterprise, or Console account" and that the free plan does not include access — this contradicts a summary reading of the pricing page, and we went with the documentation.