Coding agent leaderboard: the scores, with sources.

On the official Terminal-Bench 4.0 coding agent leaderboard, the top score is 58.2%, set by GPT-6 Astra running in OpenAI’s Codex, with Claude Fable 5.1 in Claude Code at 57.9%, a statistical tie. On SWE-Bench Pro’s public leaderboard, Meta’s Muse Spark 1.1 leads at 61.5%, but Scale AI has not yet added the newest Claude, GPT-6 or Gemini models.

Both were checked on September 28, 2026. Anthropic reports 66.4% on Terminal-Bench 4.0 for Claude Opus 5.5, released September 22, but that run is the vendor’s own and has not appeared on the leaderboard yet. Below are the current tables, where each number comes from, and what the scores cannot tell you. If the benchmarks themselves are new to you, start with AI benchmarks explained.

01Which AI model is best at coding right now?

No single model wins every coding benchmark in September 2026. The strongest independently measured results belong to GPT-6 Astra, first on Terminal-Bench 4.0 and tied first on DeepSWE, Claude Fable 5.1, tied with Astra on Terminal-Bench 4.0, and Claude Opus 5, tied first on DeepSWE.

The newest releases are not on any independent board yet: Claude Opus 5.5 and GPT-6 Sol and Luna all shipped on September 22. Our dated list of 2026 model releases shows what came out when.

02Terminal-Bench 4.0 leaderboard

Terminal-Bench 4.0 tests whether an agent can finish real work in a command-line environment. Each entry covers 66 tasks with five attempts each, 330 attempts in all, and most models run in their maker’s own agent, such as Codex or Claude Code. Checked on September 28, 2026, on the official leaderboard, last updated September 21. We show each model’s best run; cost per attempt is our division of the leaderboard’s total run cost by 330.

Three details change how you read this table. Grok 4.7’s run is marked partial on the leaderboard, with 324 of 330 attempts finished. Google’s models ran in the generic mini-SWE-agent harness, while the others ran in their makers’ own agents. And the gap between first and second place, 0.3 points, is far smaller than either margin.

Effort matters as much as the model. GPT-6 Astra scores 50.6% at low effort and 58.2% at max, and the full run’s cost rises from $1,557 to $3,267. Claude Fable 5.1 goes from 43.3% to 57.9% as its run cost rises from $2,359 to $6,244.

03SWE-Bench Pro scores

SWE-Bench Pro asks an agent to fix real issues in actively maintained repositories, and Scale AI keeps most tasks private to limit leakage. The first six rows come from Scale’s public leaderboard, checked on September 28, 2026. The rest are the model makers’ own reports, read on September 24 and 25, which nobody has reproduced on the board.

Scale also scores a private set of 276 tasks from startup codebases that are not public. There, Muse Spark 1.1 leads at 51.5%, ten points below its public score, and Claude Opus 4.6 follows at 47.1%.

Two vendor notes deserve weight. OpenAI marks its SWE-Bench Pro figure with the warning that labs have seen evidence of memorization on the public set. Qwen says it corrected problematic tasks and reran every model it compares on its refined version, so its 67.7% is not directly comparable to Scale’s board. Anthropic’s launch posts for Fable 5.1, Opus 5 and Opus 5.5 headline other benchmarks instead.

04DeepSWE v1.1 leaderboard

DeepSWE, built by Datacurve, uses 113 tasks written from scratch across 91 repositories in five languages, which it says keeps the solutions out of any model’s training data. Every model runs in the same mini-swe-agent harness. Checked on September 28, 2026, on the official leaderboard, updated September 22.

Open-weight models do well here. GLM-5.3 and Kimi K3 finish five points behind the leaders, at under 40% of Claude Opus 5’s cost per task.

05Is SWE-bench Verified still worth quoting?

Not for comparing frontier models. On February 23, 2026, OpenAI said it would stop evaluating on SWE-bench Verified, after an audit found flawed tests that reject correct fixes and signs that models had trained on the solutions. The official SWE-bench leaderboards have not added an entry since February 26, 2026. On the bash-only view, where every model runs in the same minimal agent, the top score is 76.8%, Claude Opus 4.5 at high effort.

Vendors still print Verified scores. Mistral reported 77.6% for Mistral Medium 3.5 in May, DeepSeek 80.6% for V4-Pro in April, and Anthropic 93.9% for the invitation-only Mythos Preview. Each ran its own harness, so these numbers compare poorly with each other and not at all with the frozen leaderboard. Stanford’s 2026 AI Index sums up the problem: scores on the benchmark rose from 60% to near 100% in a single year.

06Which benchmark versions are current?

Version numbers matter, because the same model can score 89% on one Terminal-Bench and 19% on another. As of September 28, 2026:

Older versions live on in launch posts. Google’s Gemini 3.8 Flash model card shows 89.4% on Terminal-Bench 2.1 next to 19.1% on version 4.0, and OpenAI’s GPT-5.6 post reports 88.8% for GPT-5.6 Sol on version 2.1. Terminal-Bench 3.0, released July 30, was replaced a month later.

Launch posts also lean on newer tests such as CursorBench 4.0 and FrontierCode 1.1. Unless a public leaderboard lists the run, treat the number as the vendor’s own measurement.

07What the scores do and do not tell you

A leaderboard score measures a whole setup, not a model in isolation. Every example below comes from the tables above.

The practical rule: a gap inside the error bars is a tie, and a vendor score is a claim until an independent board reproduces it. “Effort” is how long a model may reason before it answers, which our guide to reasoning models explains.

What no public leaderboard can tell you is how an agent does on your code. Twenty tasks from last month’s work, graded the same way for every agent, will settle more than any table here. Our guide to evals for beginners shows how to build that set, and our comparison of the best AI coding agents covers the tools, prices and plans behind these models.

08FAQ

Read next: AI benchmarks explained for what each test measures, or Claude Code vs Codex vs Cursor to pick between the agents behind these scores.