Claude Code with Opus 5.5 tops the new Coding Agent Index. What each setup costs

What the index measures

Artificial Analysis published version 2.0 of its Coding Agent Index on October 22, 2026. It doesn’t score models on their own. In the words of its announcement, it “evaluates models together with their agent harnesses,” the program that drives the model, such as Claude Code, Codex or Antigravity. The same model in another agent, or at another effort level, gets its own row.

It combines three benchmarks, weighted equally, for 290 tasks in all:

• FrontierCode v1.1, from Cognition: 100 software engineering tasks on real open-source repositories, “graded on whether a maintainer would merge the change.” Its test set is private.
• Terminal-Bench 4.0, from the Laude Institute: 66 tasks worked in a terminal.
• SWE-Atlas-QnA-AA: Scale AI’s 124 questions about code repositories, with Artificial Analysis’s revised grading and GPT-6 Luna as the judge.

Each task is tried three times and the scores averaged. An attempt that errors, runs out of time or is blocked by a safety refusal scores zero. Cost is the “average pay per token API cost per task, based on provider token pricing rather than consumer plans,” as the methodology puts it.

Version 2.0 swaps DeepSWE for FrontierCode and regrades the repository questions, which lowered that benchmark’s average by 1.7 points across the 21 setups tested on both versions. Scores from version 1.5 aren’t comparable.

Who leads

These are the ten highest-scoring setups that use a single model, as the index stood on October 28, with what a task costs at API prices and how long one took on average:

Agent and model	Index	Cost per task	Time per task	Where you get it
Claude Code, Opus 5.5 (max)	61.7	$11.42	61 min	Claude Pro or Max
Antigravity CLI, Gemini 4 Carbon (high)	59.4	$7.74	30 min	Google AI Ultra or Enterprise, since November 10
Claude Code, Sonnet 5.5 (max)	59.3	$13.30	80 min	Claude Pro or Max
Muse Code, Muse Wonder Full (max)	58.9	Not given	35 min	Muse Code Power or Maximum
Claude Code, Fable 5.1 (max, with fallback)	57.6	$11.53	35 min	Claude Max, or Pro with usage credits
Codex, GPT-6 Astra (max)	54.7	$7.04	25 min	ChatGPT Plus or higher
Codex, GPT-6.1 Sol (max)	54.1	$1.50	24 min	ChatGPT Plus or higher
Antigravity CLI, Gemini 4 Argon (high)	53.2	$11.73	35 min	Google AI Pro (paid), Ultra or Enterprise
Muse Code, Muse Wonder (max)	52.0	$8.05	25 min	Every Muse Code plan
OpenCode, GLM-5.3 (max)	48.9	$5.23	56 min	Your OpenCode provider

The plans come from Claude’s pricing, Codex pricing, Antigravity’s models page, Meta’s Muse Code page and Google’s announcement of Carbon. Artificial Analysis gives no cost for Muse Wonder Full. It prices both Gemini 4 models at $4 per million input tokens and $20 per million output, which is what Argon costs on the API from December 1, 2026; until November 30 it’s $2 and $10, by the Gemini API’s pricing page. When this memo was first published, Artificial Analysis charted Argon at its launch price, $5.86 a task.

Below the top ten, Mistral Vibe with Mistral Large 4, Mistral’s new open-weight model, joined on October 29 and scores 48.3 at high effort for $1.26 a task; at medium effort it scores 35.4. Codex with GPT-6.1 Terra, added after its release on October 26, scores 48.2 at extra high effort for $0.69 a task, and Codex with DeepSeek V4.1 Flash scores 47.3 at $0.39. Grok Build with Grok 4.7 scores 46.6 at $9.48 a task, and Kimi Code CLI with Kimi K3 41.4. Cursor isn’t on the index.

The announcement’s ranking passes over the page’s setups that use more than one model. In Cognition’s Devin Fusion CLI, the best, Claude Opus 5.5 with GPT-6 Luna, scores 61.5 at $8.00 a task: 0.2 points behind Claude Code with Opus 5.5, for 30% less.

Copilot Fusion and Gemini 4 Carbon join

The strongest setups added since October 23 are two announced on October 28.

GitHub Copilot with Copilot Fusion at Max scores 59.9, at $5.59 a task and 36 minutes. Fusion isn’t one model. GitHub’s announcement says you “choose Copilot Fusion and let Copilot intelligently route and orchestrate which models it uses for you,” sometimes with one model drafting and another reviewing, and Artificial Analysis lists Anthropic and OpenAI as the makers of the models it used. GitHub’s claim that it came “within two points of Claude Code with Opus 5.5 Max” at “about half the average API cost per task” matches the page: 1.8 points behind, at 49% of the cost, in 41% less time. It also leads Terminal-Bench, at 71.2%.

Fusion is on every Copilot plan, but GitHub’s docs put its High and Max levels on Pro+ ($39 a month), Max, Business and Enterprise only. You pay for each model it uses at that model’s standard rate, and “through November 30, 2026, users on paid Copilot plans receive a 30% discount on model costs when using Copilot Fusion.”

Antigravity’s CLI with Gemini 4 Carbon at high effort scores 59.4, at $7.74 a task and 30 minutes: 6.2 points above Argon in the same agent, for about two-thirds of the cost. Google announced Carbon the same day, when only a subset of cyber defenders could use it, and Artificial Analysis marked it as unavailable. On November 10, Google opened it to Google AI Ultra subscribers in Antigravity, and the index now lists it as available (our memo on Carbon).

What a task costs

In the October 22 announcement, “four configurations sit on the cost frontier” of score against price: Codex with DeepSeek V4.1 Flash at 47.3 for $0.39, Codex with GPT-6.1 Sol at 54.1 for $1.50, Codex with GPT-6 Astra at 54.7 for $7.04, and Claude Code with Opus 5.5 at 61.7 for $11.42.

A few things stand out:

• Copilot Fusion now beats Codex with GPT-6 Astra on both counts. It scores 5.2 points more, for $1.45 less a task.
• GPT-6.1 Sol comes within 0.6 points of GPT-6 Astra for about a fifth of the cost. OpenAI’s API prices for Sol are $2 per million input tokens and $10 per million output, against $10 and $50 for Astra.
• Opus 5.5 is 7.6 points ahead of GPT-6.1 Sol, at more than seven times the cost per task. Codex with GPT-6 Astra sits between them, at $7.04.
• Sonnet 5.5 at max cost more than Opus 5.5 at max. It came to $13.30 a task against $11.42, scored lower, and took 80 minutes against 61.
• The cheapest setup scores low. Codex with GPT-6 Luna costs $0.16 a task and scores 31.1. DeepSeek’s own harness, with the same model as the frontier’s cheapest setup, scores 46.3 at $0.30.

On a plan there’s no bill per task, but there are limits. OpenAI’s Codex pricing page gives ChatGPT Plus 15 to 160 local messages every five hours with GPT-6.1 Sol, against 5 to 45 with GPT-6 Astra.

What the scores don’t tell you

• Your effort level. The Claude Code and Codex setups near the top all ran at max effort. Claude Code starts Opus 5.5 at medium, and the index has no row for it there. Anthropic’s documentation says max “may show diminishing returns and is prone to overthinking, so test before adopting it broadly.”
• The kind of work. The three benchmarks disagree. Among single-model setups, Sonnet 5.5 leads Terminal-Bench at 66.2%, ahead of Gemini 4 Carbon at 63.6% and Opus 5.5 at 63.1%, and Copilot Fusion beats all three at 71.2%. Muse Wonder Full leads the repository questions at 68.8%. Codex with DeepSeek V4.1 Flash is second on FrontierCode among single-model setups, at 52.6%, but scores 28.3% on Terminal-Bench.
• Who wrote the tests. FrontierCode’s private test set comes from Cognition, which also makes Devin Fusion CLI.
• Fallbacks. The Fable 5.1 row runs with a “Default Fallback”: when a safety refusal blocks Fable, the agent switches to another model, “often a less capable one,” and the attempt is scored like any other. The page names Claude Opus 4.8 and Claude Opus 5 as the fallback models.
• Routers change. Copilot Fusion’s score is for the models it had on the day. GitHub’s docs say “the model pool changes as new models become available and as GitHub evaluates which models perform best.”
• How far a plan goes. The index prices a task. It doesn’t measure how many tasks a week a plan allows.

How to choose

1. For the hardest work, Claude Code with Opus 5.5. It leads the index and FrontierCode, and comes with Claude Pro ($20 a month) and Max (from $100). Keep max effort for the problems that need it.
2. For the most per dollar, Codex with GPT-6.1 Sol. It scores within 0.6 points of GPT-6 Astra at about a fifth of the API cost, and on ChatGPT Plus ($20 a month) it allows about three times as many messages every five hours.
3. If you pay per token, Codex with DeepSeek V4.1 Flash. At $0.39 a task, it’s the cheapest setup on the cost frontier. DeepSeek’s guide sets Codex up with a DeepSeek API key in place of a ChatGPT sign-in, and Codex keeps those sessions apart from your ChatGPT ones.
4. On GitHub Copilot, Copilot Fusion at Max. It scores 59.9 at about half Opus 5.5’s cost per task, and needs Copilot Pro+ ($39 a month) or higher.
5. On a Google plan, Antigravity with Gemini 4 Argon, or Carbon on Ultra. Argon scores 53.2 at high effort, and needs a paid Pro plan, or a trial on an eligible premium device. Carbon, at 59.4, has come with Google AI Ultra since November 10, and Google hasn’t said whether it will come to Pro.
6. On Muse Code, check your plan. Muse Wonder Full, fourth among single-model setups, comes with Power ($50 a month) and Maximum ($200). Muse Wonder, on every plan from $5, scores 52.0.

Using more than one

The index rates one setup at a time, but its spread of cost argues for mixing them: a cheaper setup or a lower effort for routine changes, and Opus 5.5 for the hard ones. Enjoy runs six of the index’s agents, Claude Code, Codex, Antigravity, Muse Code, Grok Build and OpenCode, side by side on your computer, each signed in with the account you already have. You can switch agents in the middle of a conversation, and the new agent gets the conversation so far. You can set the effort in your message, such as “quick one, low effort” or “think as hard as you can.” And you can see what each Claude and ChatGPT account has left.

Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t add Opus to a ChatGPT plan or GPT to a Claude one: each agent works on its own lab’s plan.

We’ll update this memo when Artificial Analysis adds the agents it says are coming, or changes the index again.