Two Claude Models Top the Coding Agent Leaderboard at 58 Points, But a Task Costs Anywhere from $1.51 to $44.22
Artificial Analysis has updated its coding agent index, which tests not the model alone but the combination of a model with a harness such as Claude Code. Claude Fable 5.2 and Claude Opus 5.5 share first place with 58 points on the index scale. Among configurations scoring 48 points or more, the cost per task varies about 30-fold: $1.51 for Codex with GPT-6.1 Sol versus $44.22 for Claude Code with Fable 5.2 (max). Choose an agent based on the pair of score and cost per task.
AI-processed from X: @ArtificialAnlys; edited by Hamidun NewsWhen a team chooses an AI programming assistant, the model name alone tells them little. The same model performs differently depending on the harness, the program that gives it access to the repository and terminal: Claude Code, Codex, Antigravity CLI, or Devin Fusion CLI. Independent lab Artificial Analysis measures exactly these combinations in its Coding Agent Index. Version 2.1 is out: Claude Fable 5.2 and Claude Opus 5.5 share first place with 58 points on the index scale, while the cost of a single task for strong setups varies roughly 30-fold, from $1.51 to $44.22.
What changed in the index
In version 2.1, two of the three tests were updated. The first is FrontierCode: version 1.1 has been replaced by FrontierCode-AA 2.0, Artificial Analysis's own implementation of the FrontierCode 2.0 suite from Cognition. It contains 132 real-world tasks, and it does not check for formal test passing. Instead it applies a criterion every developer understands: would the repository maintainer accept the change the agent proposed? It works like a code review in a real project.
The second is Terminal-Bench: version 4.0 has been replaced by 4.1. There are still 66 tasks, but the verification scripts are now more reliable and better protected against reward hacking. This term describes a situation where an agent finds a loophole and gets credit without actually solving the task. The third test, SWE-Atlas-QnA-AA, was not removed from the index.
Because FrontierCode-AA 2.0 is stricter, scores in v2.1 are lower: models lost 5 points on average. The only configuration that barely dropped is Codex with GPT-6.1 Terra (max), whose result stayed within 1 point of its previous score. For this reason, v2.1 figures cannot be compared directly with those of earlier versions, because the task set itself has changed.
Who leads
Two setups built on Anthropic models share first place with 58 points. Claude Fable 5.2 scores 58 in both xhigh and high modes in Claude Code. Claude Opus 5.5 (max) also scores 58 in Claude Code, and additionally in a pairing with GPT-6 Luna High in Cognition's multi-model environment, Devin Fusion CLI.
One detail stands out: Claude Fable 5.2 in max mode scored only 56. According to Artificial Analysis, more of its attempts ran into the per-task time limit. So the most thorough mode does not guarantee the best result: the agent may simply fail to finish in time. Next come Claude Sonnet 5.5 (max) with 56 points in Claude Code and Gemini 4 Carbon (high) with 55 points in Antigravity CLI.
Cost versus quality
For a budget, the most important figure is cost. Among configurations scoring 48 points or more, the cost per task differs by roughly 30 times. The cheapest is Codex with GPT-6.1 Sol (max): 52 points at $1.51 per task. The most expensive is Claude Code with Claude Fable 5.2 (max): 56 points at $44.22.
The quality gap between these two options is 4 points, while the cost gap is nearly 30-fold. If an agent runs hundreds of tasks a day, this is no longer an academic nuance but a major line item in the budget.
What to do in practice
First, do not transfer someone else's ranking to your own stack without checking: change the harness and the result changes. Second, look at the score together with cost and time limits: the "maximum" mode may cost more and still lose to a lighter one. Third, keep the limitations in mind: this is one independent index built from three tests, and it knows nothing about your specific codebase.
A sensible plan: pick two or three configurations from the top of the table, be sure to include one inexpensive option, and run them on your own tasks from your own repository. The public index can tell you where to start, but the cost of an accepted patch in your project should make the final call.
- Nov 21, 2026OpenAI Names 4 Reasons GPT-6 Astra Hides Its Reasoning from Monitors
- Nov 21, 2026Claude Code Taught to Hunt Vulnerabilities: 83 Skills Built from 681 HackerOne Reports
- Nov 21, 2026Anthropic: Claude responds appropriately to self-harm queries in 94% of app conversations and 60% via API
- Nov 21, 2026agweb browser for Claude Code and Codex doesn't steal focus and uses 590 MB of memory instead of 1.6 GB
Key Questions Answered
What changed in the index?
In version 2.1, two of the three tests were updated. The first is FrontierCode: version 1.1 has been replaced by FrontierCode-AA 2.0, Artificial Analysis's own implementation of the FrontierCode 2.0 suite from Cognition. It contains 132 real-world tasks, and it does not check for formal test passing. Instead it applies a criterion every developer understands: would the repository maintainer accept the change the agent proposed? It works like a code review in a real project.
Who leads?
Two setups built on Anthropic models share first place with 58 points. Claude Fable 5.2 scores 58 in both xhigh and high modes in Claude Code. Claude Opus 5.5 (max) also scores 58 in Claude Code, and additionally in a pairing with GPT-6 Luna High in Cognition's multi-model environment, Devin Fusion CLI. One detail stands out: Claude Fable 5.2 in max mode scored only 56. According to Artificial Analysis, more of its attempts ran into the per-task time limit.
What are the key details on Cost versus quality?
For a budget, the most important figure is cost. Among configurations scoring 48 points or more, the cost per task differs by roughly 30 times. The cheapest is Codex with GPT-6.1 Sol (max): 52 points at $1.51 per task. The most expensive is Claude Code with Claude Fable 5.2 (max): 56 points at $44.22. The quality gap between these two options is 4 points, while the cost gap is nearly 30-fold. If an agent runs hundreds of tasks a day, this is no longer an academic nuance but a major line item in the budget.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.