Ten AI coding harnesses, ranked S to F

Everyone argues about which model is smartest. The harness is what decides whether any of that intelligence reaches your codebase, and the gap between the best harness and the worst is bigger than the gap between the models.

There is a benchmark leaderboard that settles a lot of this, and almost nobody reads it properly.

The argument

Terminal-Bench does not rank models. It ranks pairs, each one an agent harness plus the model driving it. Once you see the results that way, the whole question of which model is best starts to look like the wrong question.

Here is one harness, Claude Code, on the same benchmark and the same version.

• With Fable 5: 83.8 percent
• With Sonnet 5: 74.6 percent
• With GLM-5.1: 58.7 percent

A 25 point spread from the model alone. Now the other direction. Claude Code with Sonnet 5 scores 74.6 percent. Terminus 2 with Gemini 3 Pro scores 73.9 percent. A mid-tier model in a good harness beats a frontier model in a worse one.

That is the whole piece. The model sets your ceiling. The harness decides how much of that ceiling you ever touch.

A note on the numbers, because most comparisons get this wrong

Nearly every comparison article we found put Codex's Terminal-Bench score against Claude Code's and declared a winner. Most of them were comparing a 2.0 score against a 2.1 score. Version 2.1 revised 28 of the 89 tasks. Those versions are not interchangeable, and a comparison across them is not a comparison.

Read the same version and the picture changes. Here is the top of the Terminal-Bench 2.1 leaderboard.

• Claude Code with Fable 5: 83.8 percent, plus or minus 1.2
• Codex with GPT-5.5: 83.1 percent, plus or minus 1.1
• Terminus 2 with Fable 5: 80.4 percent, plus or minus 1.2
• Cursor CLI with Grok 4.5: 79.3 percent, plus or minus 1.5

So the ranking below is not a benchmark ranking, because the benchmark does not separate the top two. It is a judgement on product quality, cost, safety and what the tool does to your working day, with the benchmark used where it is genuinely decisive and ignored where it is not.

S tier: Claude Code

Top of Terminal-Bench 2.1 at 83.8 percent, and more importantly the harness that holds up best when the model underneath it changes. It runs in the terminal, it edits files directly, and it asks before doing anything it cannot undo.

The reason it takes S over a statistically tied Codex is not the 0.7 points. It is that the surrounding machinery, the subagents and hooks and skills and permission modes, is where the actual work of controlling an agent happens, and it is the most complete implementation of that we have used. The benchmark measures the engine. S tier is about the car.

A tier: Codex

83.1 percent on the same benchmark and the same version. On raw capability there is nothing between it and the S tier.

What keeps it off S is single-vendor lock-in. Codex runs OpenAI models. When the frontier moves, and over the last two years it has moved roughly every four months, a harness tied to one lab moves only when that lab does. The leaderboard makes the cost of that concrete. The same harness scores 83.1 percent with GPT-5.5 and 78.4 percent with GPT-5.6 Terra. If you cannot swap the model, you inherit whatever your vendor ships next.

A tier: Cursor

Cursor earns A on an argument nobody else on this list can make. It is the only one that made the economics interesting.

Composer 2.5, its in-house model, scored 62 on Artificial Analysis's Coding Agent Index. That is third place, behind Claude Opus 4.7 at 66 and GPT-5.5 at 65. It costs 7 cents per task. The two models above it cost 4.10 dollars and 4.82 dollars. That is roughly 60 times cheaper for about 94 percent of the score.

For an editor-first workflow where a human reviews every diff, that trade is obviously correct, and it is why Cursor is A rather than B. It is not S because it is an editor with an agent in it rather than an agent with an editor attached, and on long autonomous runs that distinction shows.

B tier: Antigravity

Google's agent-first IDE, and the designated survivor of Gemini CLI, which stopped serving Pro, Ultra and free accounts on 18 June 2026 and pointed everyone here.

B is a distribution ranking rather than a quality one. It is free to start, it is Google's, and it will end up in front of an enormous number of developers who never evaluate an alternative. What holds it at B is that the free tier has been cut repeatedly since launch and Google has not committed to keeping one. Building a team's workflow on a metered preview with no pricing guarantee is a real risk.

B tier: Cline

Open source, bring your own key, and an approval step at each stage rather than a long autonomous run you audit afterwards.

That approval step is the entire argument for it. Cline is the harness to hand someone who does not yet trust agents, because it cannot surprise you. The cost is throughput. A tool that stops and asks is a tool you have to sit with, and a step you approve every time is a step you will eventually stop reading. B is right. Genuinely good, and deliberately slower.

C tier: OpenCode

MIT licensed, more than 75 providers, and a terminal agent you can point at essentially any model including one running on your own hardware.

C is not a criticism of the engineering, which is good. It is that OpenCode's quality is almost entirely a function of the model you choose, and the project cannot control that. Look at the spread again. The same harness with a weaker model loses 25 points. A tool whose result depends that heavily on a decision it hands back to you belongs in the middle of the table. If you know exactly which model you want and why, move it up two tiers.

C tier: Aider

The original terminal coding agent, and the one that established the pattern everything above it now uses. It deserves the credit.

It sits in C because the last push to the repository was 22 May 2026, three months before this was written, and the project's own leaderboard has not been refreshed for 2026 models. The repository is not archived and the licence is permissive, so nothing stops you using it. But a harness's job is to keep pace with the models, and this one has stopped. In a category where the frontier moves quarterly, three months of silence is a tier.

C tier: OpenHands

Genuinely good at what it is actually for, which is building and running agent workflows. It is a research platform and a framework, MIT licensed and actively developed.

That is also why it is C. If your goal is to close a ticket this afternoon, a framework for constructing agents is not the shortest path. It is an excellent tool for the problem of how agents should work, and a mediocre one for the problem of please fix this bug. Rank is against the job, and the job here is shipping.

D tier: Devin

Devin is the most improved tool on this list and still a D.

Cognition's own 2025 review reports that 67 percent of Devin's pull requests are now merged, up from 34 percent. That is a doubling in a year and a genuinely impressive engineering result. It is also, read plainly, a statement that one pull request in three is thrown away.

The same review is candid about why. Devin does best with clear requirements, struggles with ambiguity, and usually performs worse when you keep telling it more after it has started. That is a precise description of most real software work. Devin is positioned as a junior engineer you delegate to, and it is weakest exactly where delegation is supposed to save you effort, which is the briefing.

For the record, Windsurf is not a separate entry here because it is not a separate product. Cognition rebranded it Devin Desktop on 2 June 2026 and the old Cascade agent reached end of life on 1 July. Any comparison still listing both is out of date.

F tier: Microsoft Copilot

And here is the one that matters, because it is the one you already have.

Microsoft Copilot, the assistant in your Microsoft 365 login next to Word and Teams, is not a coding harness. It has no agent loop, it does not hold your repository, it cannot run your tests, and it cannot edit a file and then check whether it broke something else. It is a chat window that returns code you paste somewhere else.

It is in F not because it is bad software. It is in F because of what it costs you to find that out. Every other tool on this list requires a decision, whether that is an evaluation, a card, an install or a conversation with someone. Microsoft Copilot requires none. It is already provisioned, already paid for, and already sitting in the tab next to your email. So it is the one a business tries first, and the one that quietly sets the expectation that AI coding tools are not much use.

What to do with this

If you take one thing from it, stop choosing the model and start choosing the harness. The model is a setting inside the harness, and on this evidence you can change it in an afternoon. The harness is where your permissions, your review step, your context and your team's habits actually live, and changing that is a quarter.

And if the tool your team currently uses for AI coding came bundled with your Office licence, you have not evaluated anything yet.