Claude Code With Opus 5.5 Tops the New Coding Agent Index at $11.42 a Task. DeepSeek's Flash Is Second on the Hardest Test for $0.39.

Artificial Analysis replaced its Coding Agent Index today, October 22, 2026. Version 2.0 is an equal-weight average of three tests: FrontierCode v1.1 (100 private real-repository tasks built by Cognition), Terminal-Bench 4.0 (66 tasks), and SWE-Atlas-QnA-AA (124 questions from Scale AI, graded by AA with GPT-6 Luna at xhigh as the judge, five votes each). FrontierCode replaces DeepSWE from v1. The leaderboard is a list of harness-plus-model configurations, not models, which is the right way to measure coding agents and the wrong way to read it quickly.

The top line is simple: Claude Code running Opus 5.5 at max leads at 61.7, and each of its tasks costs $11.42. The line underneath it is the one worth a post. Codex running DeepSeek V4.1 Flash at max scores 47.3 overall, ninth in our table, and scores 52.6% on FrontierCode, second only to the leader on the hardest of the three tests, at $0.39 per task. One is 29 times the price of the other. We read AA's live table, the methodology page, and the two secondary pieces that circulated. Here is what held up.

The Table, Read From AA's Page

Three things the table says that the headline does not:

• Sonnet costs more than Opus here. Claude Code with Sonnet 5.5 bills $13.30 per task against $11.42 for Opus 5.5, for a lower score. Sonnet 5.5's per-token price is half of Opus 5.5's; its token appetite at max effort, which we covered when it launched, is what reverses the order.
• Harness matters as much as model. The same DeepSeek Flash scores 52.6% on FrontierCode inside Codex and 46.5% inside DeepSeek's own harness. Fifteen of the 24 configurations on the board land between 43% and 49% on FrontierCode, by our count from the live table; the harness is a bigger lever than most of the model swaps.
• Muse Wonder Full has no cost. AA's table carries no dollar figure for Meta's model, which has no public API. Do not let anyone fill that cell for you.

The Benchmarks, Sorted by Who Built Them

FrontierCode is private, and its builder sells an agent on this board. FrontierCode v1.1 is 100 tasks on real repositories that are not public, built by Cognition. Cognition also makes Devin, and the Devin Fusion CLI configuration sits second at 61.5. AA's methodology discloses the authorship. We are not alleging anything; we are noting that one third of the index is a closed test owned by a company whose product is being ranked by it, and that readers should know that before quoting a FrontierCode number as neutral ground.

SWE-Atlas is judged by an OpenAI model. The question-answering third uses GPT-6 Luna at xhigh as the grader with five votes. Codex configurations are on the same board. The same disclosure applies.

Terminal-Bench 4.0 is the public, reproducible third. It is also where DeepSeek Flash tanks, at 28.3%, which is how a model that is second on FrontierCode ends up ninth overall.

v1 and v2 are not comparable. AlphaSignal reports that the mean score of configurations shared between the two versions fell from 54.8 to 47.4, with 20 of 21 lower. That figure is AlphaSignal's, not a headline on AA's page, so we print it as theirs. The direction is not in doubt: v2 dropped DeepSWE, added a harder private set, and most numbers went down. Do not put a v1 score next to a v2 score.

What the DeepSeek Line Means

FrontierCode is the test AA describes as the hardest of the three, and Codex running a model that lists at $0.15 per million tokens off-peak scores 52.6% on it, above Opus 5.5 in Devin Fusion, above Sonnet 5.5, Fable 5.1, Astra, and GPT-6.1 Sol. Then it scores 28.3% on Terminal-Bench 4.0 and the average collapses. The honest summary is that cheap Flash is a real contender on one kind of coding task and a weak one on another, and anyone arguing "cheap Flash" should quote the FrontierCode score and the cost per task separately from the index.

For readers keeping score on the model side: on AA's Intelligence Index, DeepSeek Flash 1017 sits at 45, tied with GLM-5.3, one point under MiMo-V2.6-Pro at 46.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use the index as a harness-and-model shopping list, not a model ranking. If you run Claude Code, Opus 5.5 at max is the top configuration and Sonnet 5.5 at max is not the cheaper one here. If you run Codex, GPT-6.1 Sol at $1.50 a task is the value line, and DeepSeek Flash at $0.39 is worth a trial on repository-style work.

Skip the overall number when you argue about any single model. Quote the component, the harness, and the cost; the average hides all three. And re-pull AA's page the morning you cite it; the table is live and the board moved three times last week.

The frontier is elsewhere: the top of this board is the same lab that tops the Intelligence Index, in its own harness, at $11 a task. The news is at the bottom, where a $0.39 task is second on the hardest test. Our model comparison has the current standings side by side.

Sources

• Artificial Analysis: Coding Agent Index
• Artificial Analysis: coding agents benchmarking methodology
• AlphaSignal: Artificial Analysis ranks Claude Opus 5.5 first in tougher Coding Agent Index
• DigIntoAI: Coding Agent Index v2 and FrontierCode

*Not sure which harness and model pair fits your repository? Take the free 2-minute quiz and get matched. Then read about the $0.39 model at the bottom of this table and the same-name swap that got it there: DeepSeek V4.1 Flash 1017.*