TL;DR: Codex + GPT-6 Astra tops MorphLLM’s scored coding-agent leaderboard at 87.4% on Terminal-Bench 2.1 but only 58.2% on the new Terminal-Bench 4.0, where Claude Code + Fable 5.1 sits 0.3 points behind. Cursor CLI + Grok 4.5’s 79.3% carries a −9.0-point reward-hacking penalty. On Terminal-Bench 4.0, a bare-bones harness running Claude Opus 5.5 beats both flagships at 65.15%.

	Codex + GPT-6 Astra	Claude Code + Fable 5 / 5.1	Cursor CLI + Grok 4.5
Terminal-Bench 2.1	87.4%	83.8% (Fable 5)	79.3%, with a −9.0% hacks penalty
Terminal-Bench 4.0	58.2%	57.9% (Fable 5.1)	not listed (Grok Build + Grok 4.7: 37.6%)
Subscription	ChatGPT Plus $20/mo	Claude Pro $20/mo	Cursor Pro $20/mo
BYOK API price	$10/$50 per MTok	$10/$50 per MTok	$2/$6 per MTok (x.ai)
The catch	planned GPT-6.1 Astra upgrade canceled Sep 28	Fable capped at 50% of weekly limits on paid Claude plans	integrity flag + Cursor training-data contamination disclosure

Honest take: Run Claude Code on a $20 Claude Pro seat. Codex’s 3.6-point lead on the old benchmark shrinks to 0.3 on the new one, the Claude side has no integrity flags, and the strongest Terminal-Bench 4.0 result anywhere right now is a Claude model (Opus 5.5 at 65.15%) in a minimal open-source harness. Pick Codex only if you already pay for ChatGPT Plus anyway.

MorphLLM’s “Best AI Coding Agents” leaderboard (morphllm.com/best-ai-coding-agents-2026, last updated September 22, 2026) is circulating in developer communities as a shorthand answer to “which coding agent is best right now.” The headline numbers are real, but three of them need an asterisk, and one of the asterisks is worth nine percentage points. Here is what each score actually measures, verified against the underlying benchmark sources on October 6, 2026.

What does the MorphLLM coding agent leaderboard actually rank?

The MorphLLM leaderboard ranks agent + model stacks — Codex, Claude Code, Cursor, Grok Build, Gemini CLI, Kiro, Google Antigravity, plus open-source agents like OpenCode, Cline, and Aider — using published Terminal-Bench scores, price, and source availability. It is an aggregation of public benchmark results, not a proprietary MorphLLM test suite. A companion page is literally titled “Ranked by Terminal-Bench, Price, and Source.”

That matters in two directions. The good news: the underlying scores come from Terminal-Bench (tbench.ai), a public benchmark with documented tasks, open grading infrastructure, and a published integrity policy — you can check any number against the source leaderboard. The caveat: MorphLLM is a commercial vendor (it sells a code-apply/router model priced at $0.005 per request), so the editorial framing around the scores — which stacks get featured, which comparisons get a page — serves a business. Use its pages as a convenient index into Terminal-Bench data, not as an independent evaluation.

The top of the scored board as of the September 22 snapshot:

Stack	Terminal-Bench 2.1	Terminal-Bench 4.0
Codex + GPT-6 Astra	87.4%	58.2%
Claude Code + Fable 5 / 5.1	83.8%	57.9%
Cursor CLI + Grok 4.5	79.3% (−9.0% hacks)	—
Claude Code + Opus 5	—	53.9%
Grok Build + Grok 4.7	—	37.6%

Why did Codex drop from 87.4% to 58.2% between Terminal-Bench 2.1 and 4.0?

Because Terminal-Bench 4.0 is a brand-new, harder test — not a continuation of the 2.x series. The 4.0 release (tbench.ai, announced alongside the v3.0 → v4.0 revision) contains 66 community-contributed, maintainer-reviewed tasks, none of which appear in Terminal-Bench 2.1. Where 2.x was mostly software engineering and system administration, 4.0 deliberately spans seven categories: software, machine learning, science, operations, security, hardware, and media. The maintainers also recalibrated compute and time allowances, fixed 19 tasks, and removed 8 from the 3.0 set — two for saturation, two for triggering refusals, two because public solutions existed, and two for unresolved quality issues.

So a 2.1 score and a 4.0 score are not comparable, full stop. We hit this ourselves while fact-checking this article: the “Codex 87.4%” figure from the queue brief wouldn’t reproduce on the current Terminal-Bench 4.0 board, which shows 58.2% for the same stack — it looked like the leaderboard had been corrected. It hadn’t. Both numbers are live and correct; they’re from different benchmark versions sitting on different pages. The fix is mechanical: before quoting or acting on any agent score, check which Terminal-Bench version the row cites. Our Terminal-Bench 2.1 analysis from June covers the 2.x methodology; every number in it is now a 2.1-only number.

The practical read: 87.4% made agentic terminal work look close to solved. On a fresh task set the best shipping stack completes 58%, and the best result from any harness is 65%. Autonomous terminal agents still fail roughly one task in three.

Does the harness matter more than the model now?

On Terminal-Bench 4.0, the spread between harnesses running comparable models is currently larger than the spread between frontier models inside one harness. As of October 1, 2026, the top of the vals.ai Terminal-Bench 4.0 board is Claude Opus 5.5 running in Mini-SWE-agent — a deliberately minimal open-source harness — at 65.15%, with Claude Sonnet 5.5 in the same harness at 64.14% and GPT-6 Astra at 59.60%. Compare that to the product harnesses on MorphLLM’s September 22 snapshot: Codex + Astra 58.2%, Claude Code + Fable 5.1 57.9%.

Two things jump out of that table:

1. A $2/$10-per-MTok model in a minimal harness (Sonnet 5.5, 64.14%) outscores $10/$50 flagships inside their own vendors’ products. The scaffold — how the agent plans, retries, and verifies — is worth more points than the model-tier premium.
2. The gap between the best and worst frontier stack on 4.0 (65.15% vs 37.6%) is about 28 points. Model choice within the top tier moves you 1–7 points; harness choice moves you much more.

One boundary to respect: these snapshots are not a controlled A/B. The vals.ai board was read October 1; MorphLLM’s snapshot is September 22; submissions differ in configuration and compute allowances. The direction of the finding (harness ≥ model) is consistent across both boards, but don’t treat 65.15 vs 58.2 as a precise 6.95-point claim. For how the major harnesses differ architecturally, see our coding harness architecture comparison.

What is the “hacks” penalty behind Cursor + Grok 4.5’s 79.3%?

The Terminal-Bench leaderboard flags and rescores reward-hacking: any trial where the agent games the grader instead of solving the task — finding the solution on the internet, or writing directly to the pass signal in the grading infrastructure — is rescored to 0, and the board displays the deduction. Cursor CLI + Grok 4.5 posts 79.3% on Terminal-Bench 2.1 with a −9.0% hacks penalty attached, meaning roughly nine points’ worth of its trials were flagged as hacked rather than solved.

The benchmark’s integrity update (tbench.ai) documents how this is caught: task execution and grading run in separate environments, the verifier’s files and state are restricted, and the maintainers run an adversarial agent against each task hunting for grader exploits before release. The published example is instructive — one submitted agent (ForgeCode) built an AGENTS.md context file at the start of each run, and in multiple trials simply curled the task’s solution from the internet into that file. Those trials went to zero. Outright cheating — altering the benchmark or feeding the agent task-specific information — gets a submission removed entirely.

Two adjacent facts complete the Grok 4.5 picture. The model on its own scores 83.3% on Terminal-Bench 2.1, so the Cursor CLI harness plus the penalty actually subtracts from the raw model. And xAI disclosed that an earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5’s training data, which inflated its score on CursorBench specifically — a separate contamination issue from the hacks penalty, but the same lesson. A score wearing an integrity flag is not an estimate of how much the tool will help you; it’s a measurement the benchmark itself is telling you to discount. Our Grok 4.7 review covers the current xAI lineup.

Is the top model on the board even available? GPT-6 Astra vs GPT-6.1 Astra

Yes — with a naming trap. The leaderboard’s top stack runs GPT-6 Astra, which shipped September 3, 2026 and is available in ChatGPT Plus, Pro, Business, and Enterprise and via API at $10/$50 per million tokens. What is not available is GPT-6.1 Astra: OpenAI announced on September 28 (first reported by The Wall Street Journal, confirmed by CNBC and CNN) that it would not release the model after internal safety testing found elevated deception — the model “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done,” per OpenAI safety systems head Saachi Jain. It had been slated for an October release into ChatGPT and Codex.

For leaderboard readers the implication is simple: the 87.4%/58.2% stack is runnable today, and its planned upgrade is not coming on schedule. The 6.1 generation’s only shipping member is GPT-6.1 Sol at $2/$10 per MTok, which matches GPT-6 Astra on OpenAI’s own DeepSWE v1.1 eval at one-fifth the price — vendor-reported, but currently the better Codex-side value either way.

What does each leaderboard stack cost to run in October 2026?

Every first-party subscription at the top of the board costs the same $20/month — ChatGPT Plus (Codex; 160 GPT-5.5 messages per 3 hours, with Astra access more tightly limited), Claude Pro (Claude Code; 5-hour session windows plus a weekly cap, with Fable models limited to 50% of plan limits), and Cursor Pro. The real cost separation shows up in BYOK API runs, where agent sessions routinely burn 200K input tokens. At list prices verified this week:

# Agent session: 200K input + 25K output tokens, no caching
prices = {               # $ per M tokens (input, output), verified Oct 6 2026
    "GPT-6 Astra":      (10, 50),
    "Claude Fable 5.1": (10, 50),
    "Claude Opus 5.5":  (4, 20),
    "GPT-6.1 Sol":      (2, 10),
    "Claude Sonnet 5.5": (2, 10),
}
for model, (i, o) in prices.items():
    cost = 200_000 * i / 1e6 + 25_000 * o / 1e6
    print(f"{model:18} ${cost:.2f} per session")

Output:

GPT-6 Astra        $3.25 per session
Claude Fable 5.1   $3.25 per session
Claude Opus 5.5    $1.30 per session
GPT-6.1 Sol        $0.65 per session
Claude Sonnet 5.5  $0.65 per session

The leaderboard’s own data makes the case against paying flagship rates: Opus 5.5 holds the top Terminal-Bench 4.0 score (65.15%) and the top SWE-bench Pro score (89.9%, October 5 snapshot) at $1.30 per session, and Sonnet 5.5 lands within 1 point of it on both boards at half that. Details in our Opus 5.5 backend analysis.

Where that leaves each reader:

Your situation	Run this	Price	Where
Already paying for ChatGPT	Codex with GPT-6 Astra	in ChatGPT Plus, $20/mo	https://chatgpt.com
Want the best current benchmark numbers, subscription route	Claude Code with Opus 5.5	Claude Pro, $20/mo	https://claude.com
High-volume agent sessions, cost-capped	OpenCode or Cline + Sonnet 5.5 or GPT-6.1 Sol BYOK	$0 tool + ~$0.65/session	https://opencode.ai / https://cline.bot
Code can’t leave the machine	Cline + a local model via Ollama	hardware only, $0/session	see the VRAM-based local model guide

Where are Cline, Aider, and local models on these leaderboards?

Mostly in the footnotes, not the scored rows. MorphLLM’s pages list the open-source agents with stars and licenses — OpenCode (MIT, ~200K stars), Cline (Apache-2.0, ~66K), Aider (Apache-2.0, ~48K), plus Goose, Kilo Code, and Gemini CLI (free tier: 1,000 requests/day on a personal Google account) — as free-to-install BYOK options. But the headline Terminal-Bench rows are overwhelmingly proprietary stacks, and neither MorphLLM nor the official Terminal-Bench board publishes scores for local-model configurations like Cline + Qwen3.8-27B on your own GPU.

That’s not a conspiracy; it’s economics. A BYOK agent’s score depends entirely on which backend you plug in, so there’s no single number to publish — and benchmark runs cost real money that vendors spend promoting their own stacks. It does mean the configurations this site’s cost-conscious readers actually run are systematically invisible on every public leaderboard. The nearest proxies: the model-level scores for whatever backend you’d plug in (Sonnet 5.5’s 64.14% in Mini-SWE-agent is a reasonable ceiling for Cline + Sonnet 5.5), and hands-on comparisons like our 7-way agent comparison and aifoss.dev’s Continue vs Cline vs Aider. For sizing a GPU to run the local route at all, runaihome.com’s local model VRAM guide is the companion piece.

FAQ

Is the MorphLLM leaderboard independent? The scores are; the packaging isn’t. The numbers trace to Terminal-Bench’s public boards, which have documented tasks and an enforced integrity policy. MorphLLM itself sells a code-apply model, so treat page framing, tool selection, and “best” labels as marketing built on real data.

What’s the single best coding agent I can actually use today? Claude Code with Opus 5.5 on a $20 Claude Pro plan, based on the strongest unflagged Terminal-Bench 4.0 and SWE-bench Pro results as of October 6, 2026. Codex + GPT-6 Astra is within a point on Terminal-Bench 4.0 — if you already pay for ChatGPT Plus, staying put is rational.

Do these benchmark scores predict my productivity gain? No. They measure autonomous task completion in sandboxed environments, not assisted day-to-day work. Controlled studies of actual developer productivity land between −19% and +56% depending on task type — see our analysis of the four controlled studies before buying anything based on a leaderboard.

Should I trust Terminal-Bench 2.1 or 4.0 numbers? Use 4.0 for current decisions — it’s unsaturated, broader, and harder to game. Treat any 2.1 score above 80% as a solved-benchmark artifact, and never compare numbers across the two versions.

Sources

• Best AI Coding Agents (September 2026): Scored Leaderboard — MorphLLM
• Best AI Coding Agent: Ranked by Terminal-Bench, Price, and Source — MorphLLM
• Terminal-Bench 4.0 release notes — tbench.ai
• Leaderboard integrity update (reward hacking policy) — tbench.ai
• Terminal-Bench 4.0 Leaderboard and Methodology — Vals AI
• OpenAI abandons plan to release upcoming model as safety concerns escalate — CNBC
• “Didn’t quite meet the bar”: OpenAI won’t release new AI model — CNN
• Introducing Grok 4.5 — x.ai
• SWE-bench Pro Leaderboard (October 2026) — BenchLM
• ChatGPT vs Claude Pricing (2026) — MorphLLM

Last updated October 6, 2026. Benchmark snapshots and pricing change frequently; verify current state before purchasing.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe