Choosing an AI Coding Agent

Which tool should I try first for the way I work, and what would make me switch? Public leaderboards can strike some candidates off the list; the rest you test on your own repository. On the October 2026 board, the gaps between the leading entries are smaller than their margins of error; what separates them is cost.

The leaders are closer than their margins of error

Terminal-Bench 4.0 gives an agent multi-step tasks in isolated terminal environments and counts how many it finishes on its own. Each leaderboard row pairs a harness with a model. The harness is the program around the model: it decides which files the model can read, which commands it can run, and how it retries after an error. The same model in a different harness gets a different score. When I checked on 2026-10-29, these were the main entries on the leaderboard:

Harness	Model (reasoning effort)	Resolution rate (95% CI)	Cost listed
Claude Code	Opus 5.5 (max)	64.8% ± 3.1	$4.7k
Claude Code	Sonnet 5.5 (max)	61.8% ± 2.9	$7.3k
Codex	GPT-6.1 Sol (max)	58.2% ± 3.1	$0.6k
Codex	GPT-6 Astra (max)	58.2% ± 2.8	$3.3k
Claude Code	Fable 5.1 (max)	57.9% ± 3.8	$6.2k
Grok Build	Grok 4.7 (xhigh)	37.6% ± 3.5	$3.7k
Muse Code	Muse Spark 1.3 (xhigh)	14.5% ± 2.9	not listed

Opus 5.5 leads at 64.8%, with a margin of 3.1 points, so its interval reaches down to 61.7%. Sonnet 5.5, in second place, reaches up to 64.7%. The two intervals overlap, so the board cannot tell these rows apart. The same holds for third to fifth place: both Codex rows and Fable 5.1 sit around 58%. What the table can safely do is strike off the lower half; which of the top five suits you, you still have to test on your own repository.

Where the ranking cannot separate entries, cost can. GPT-6.1 Sol and GPT-6 Astra resolve the same share of tasks, but cost $0.6k against$3.3k, more than five times less. That matches the gap in their list prices (see the model and API radar). Sonnet 5.5 is listed at half the price of Opus 5.5 yet costs more here: it used 19.4B tokens, more than twice the 8.0B of Opus 5.5. A model that is cheaper per token can spend the saving on extra steps.

What this board compares is each vendor's whole package. Almost every row is a vendor's own harness running its own model. The exceptions are GLM-5.3 and Qwen3.8-Max, which ran inside Claude Code and scored 41.8% and 27.0%.

Several tools have no entry. Google's models appear only as Gemini 3.8 Flash in the generic mini-SWE-agent harness, at 19.1%; Gemini CLI itself is not listed. Neither are Cursor, OpenCode, Pi or Copilot CLI. Cursor is centered on an IDE workflow that a terminal benchmark barely measures. Pi can be extended into many different harnesses, so “Pi” without its model and extensions is not one stable test subject.

Version 4.0 uses a new task set, and its scores run much lower than 2.0's. On the 2.0 leaderboard, Codex CLI with GPT-5.5, released in April, scored 82.2%. Numbers from the two versions cannot be compared.

What each tool buys

Tool	Why I would choose it	What I would watch
Claude Code	Opus 5.5 and Sonnet 5.5 lead Terminal-Bench 4.0; cohesive repository workflow with instructions, hooks, MCP, permissions, and subagents	behavior can become hard to trace after several layers of hooks and agents
Codex CLI	the lowest cost among the Terminal-Bench 4.0 top five (with GPT-6.1 Sol); sandboxing, approvals, scripting, and implement–test loops	local, scripted, and cloud modes may not share the same permission behavior
Antigravity CLI (agy)	Google's terminal entry point for individuals now: the free plan and Google AI Pro / Ultra both use it, and Gemini 4 Argon is available only there; multi-agent	the free quota refreshes weekly but has no published number (see free allowances); the original Gemini CLI now serves only Code Assist organisations
GitHub Copilot CLI	GitHub-native identity, repositories, issues, pull requests, custom agents, skills, and MCP	its strongest value depends on the GitHub and Copilot ecosystem
OpenCode	one terminal interface for several providers and local models	plugins, model aliases, and provider settings make two installations behave differently
Aider	Git-native commits, repository maps, and explicit edit and architect workflows	it is a focused pair-programming tool rather than a general programmable harness
Cursor	low-friction code search, editing, diff review, and background work inside the editor	benchmark scores say little about indexing, privacy settings, or review ergonomics
Pi	small core, multiple providers, branchable sessions, and enough extension points to build a custom harness	the user owns isolation, permissions, extension review, and many team features
Muse Code	lead agent with bounded workers, optional per-worker Git worktrees, background reviewers, and an append-only event log for resume and rewind	young product, at 14.5% on Terminal-Bench 4.0; workflows depend on build and rollout, and more agents also means more cost to coordinate and review (multi-agent coordination)

The nine tools do not need the same role. Cursor can be the best place to review a diff while Codex or Claude Code does the terminal work. Pi is useful when the experiment itself is the harness. OpenCode emphasizes provider freedom, Copilot CLI minimizes switching for GitHub-centered teams, and Aider keeps the workflow close to Git.

How I would choose today

• For terminal-first repository work, start with Codex CLI and GPT-6.1 Sol, the cheapest of the top five; hand the tasks it fails to Claude Code with Opus 5.5.
• For an IDE-centered review loop, start with Cursor.
• For Google models, or a zero-cost start, try Antigravity CLI (agy).
• For provider switching, try OpenCode.
• For a GitHub-centered workflow, try GitHub Copilot CLI.
• For Git-first terminal pairing, try Aider.
• For harness experiments and controlled model comparisons, use Pi.
• For long repository tasks split across several coordinating agents, try Muse Code on small tasks first, and check whether the extra agents actually shorten the work.

Then keep the one that fails least expensively on the actual repository. Familiarity matters: a slightly weaker tool with predictable permissions and recovery can be more useful than a benchmark leader that does not fit the workflow.

A personal replay that would change the answer

Give each serious candidate the same clean repository copy, budget, permissions, and tasks:

Task	What to record
fix an unfamiliar defect	test result, time to locate, unrelated edits
make a cross-file refactor	API preservation, diff size, rollback effort
recover after a failed attempt	repeated mistakes and second-attempt result
handle a permission trap	out-of-scope reads or writes and whether approval worked
resume an interrupted session	lost state and recovery time
run non-interactively	exit code, logs, timeout, and cost cap

Keep task completion, wall time, model/token cost, human interventions, dangerous actions, and cleanup effort. Pass rate alone hides the part that usually determines whether a tool is pleasant to use. Treat reasoning effort as a variable too: Anthropic's own Sonnet 5.5 results show that the highest setting is not automatically the best, which is why the evaluation protocol lists it separately.

My current shortlist can change quickly because all of these products move quickly. The replay protocol should change much less. That is the durable part of this note.

My reading