Field intelligence for AI-first professionalsVol. II · Nº 195 · Monday, November 16, 2026
← Phantom Notes
Nº 192AI AgentsNovember 13, 20265 min read

Claude Fable 5.2 Scores Lower on Max Than on High in Artificial Analysis's Rebuilt Coding Index, at Six Times the Cost per Task. Claude Code Also Handed 5% to 9% of Attempts to a Fallback Model After Safety Refusals.

Artificial Analysis released Coding Agent Index v2.1 on November 13, replacing FrontierCode v1.1 with FrontierCode-AA 2.0, AA's implementation of Cognition's 132-task benchmark graded on whether a maintainer would merge, and Terminal-Bench 4.0 with 4.1. Claude Code with Fable 5.2 at high or xhigh, Claude Code with Opus 5.5 at max, and Cognition's Devin Fusion CLI tie at 58. Fable 5.2 at max scores 56 at $44.22 per task against 58 at $7.51 on high, because 16% of its FrontierCode attempts and 13% of its SWE-Atlas attempts hit the time limit. Safety-refusal fallbacks completed 9% of attempts for Opus 5.5 at max and Devin Fusion, 6% for Fable 5.2 at xhigh, and 5% for Sonnet 5.5. Codex with GPT-6.1 Sol scores 52 at $1.51. Every configuration scored lower on the new FrontierCode, and v2.0 numbers are not comparable. We read AA's thread and page.

By T.W. Ghost


Artificial Analysis announced Coding Agent Index v2.1 tonight, November 13, 2026, at 6:04 PM Central. Two of the three equal-weight components changed. FrontierCode v1.1 is replaced by FrontierCode-AA 2.0, AA's own implementation of Cognition's FrontierCode 2.0: 132 private real-repository tasks graded on whether a maintainer would merge the result. Terminal-Bench 4.0 is replaced by 4.1, the same 66 tasks with updated verifiers. SWE-Atlas-QnA-AA stays. At the top, three configurations tie at 58: Claude Code with Claude Fable 5.2 at xhigh and at high, Claude Code with Opus 5.5 at max, and Cognition's own Devin Fusion CLI running Opus 5.5 at max with GPT-6 Luna at high. Next come Fable 5.2 at max and Sonnet 5.5 at max at 56, and Gemini 4 Carbon at high in Antigravity CLI at 55.

Three things in AA's thread matter more than the podium. First, more effort is not better: Fable 5.2 in Claude Code scores 53 at medium for $2.89 a task, 58 at high for $7.51, 58 at xhigh for $15.47, and 56 at max for $44.22, because at max 16% of FrontierCode attempts and 13% of SWE-Atlas attempts hit the time limit and score zero, against 1% and 0% at xhigh, while using about three times the output tokens. Second, a share of the Anthropic results were finished by someone else: when the model refused on safety grounds, a fallback model completed the attempt, 9% of them for Devin Fusion CLI and for Opus 5.5 at max, 6% for Fable 5.2 at xhigh, 5% for Sonnet 5.5 at max; Codex blocked about 1% for GPT-6.1 Sol and GPT-6 Astra at max. Third, the benchmark owner ties for first. FrontierCode is Cognition's, and Cognition's Devin Fusion CLI shares the top score, the same catch we raised when v2.0 adopted FrontierCode. We read AA's thread and its coding agents page. Here is what held up.


The Board

ConfigurationIndex v2.1Cost per taskWho
Claude Code + Fable 5.2 (xhigh)58$15.47Independent (AA)
Claude Code + Fable 5.2 (high)58$7.51Independent (AA)
Claude Code + Opus 5.5 (max)58Independent (AA)
Devin Fusion CLI (Opus 5.5 max + GPT-6 Luna high)58Independent (AA); Cognition owns FrontierCode
Claude Code + Fable 5.2 (max)56$44.22Independent (AA)
Claude Code + Sonnet 5.5 (max)56Independent (AA)
Antigravity CLI + Gemini 4 Carbon (high)55Independent (AA)
Claude Code + Fable 5.2 (medium)53$2.89Independent (AA)
Codex + GPT-6.1 Sol (max)52$1.51Independent (AA)

AA notes about a 30-fold cost range among configurations scoring 48 or higher, and that Fable 5.2's medium, high, and xhigh all sit on the Pareto line.


What Changed Underneath

DetailWho
FrontierCode-AA 2.0All 23 configurations scored on both versions came down; mean 45% to 29%; Terra at max fell under a point, Astra and Sol 7 points, Opus 5.5 12 points; index average down 5 points; Terra the only configuration not to fallIndependent (AA)
Terminal-Bench 4.1All 23 configurations moved under 3 points; Kimi Code CLI with K3.1 Pro leads at 68%, Sonnet 5.5 max 66%, Opus 5.5 max and Carbon high 65%Independent (AA)
Component scoresFable 5.2 xhigh: FrontierCode 49%, SWE-Atlas 62%, Terminal-Bench 64%; Opus 5.5 max: 44%, 66%, 65%; Devin Fusion: 44%, 68%, 63%Independent (AA)
Time limits at maxFable 5.2 max: 16% of FrontierCode and 13% of SWE-Atlas attempts score zero (xhigh: 1% and 0%); about 3x xhigh's output tokens on FrontierCodeIndependent (AA)
Safety-refusal fallbacksCompleted 9% of attempts for Devin Fusion CLI and Opus 5.5 max; 6% Fable 5.2 xhigh; 5% Sonnet 5.5 max; Codex blocks about 1% for GPT-6.1 Sol and GPT-6 Astra at maxIndependent (AA)
Two FrontierCodesCoding Agent Index FrontierCode scores differ from Intelligence Index FrontierCode, which uses AA's Stirrup harnessIndependent (AA)
Per-configuration tableJavaScript-rendered on AA's page; not scraped in our readNot rechecked

Max effort loses to high and costs about six times as much. Fifty-six for $44.22 against 58 for $7.51. The clock, not the model, is the reason, and the clock is part of the job.

Some "Claude" results were finished by another model. AA's fallback figures mean up to 9% of a configuration's attempts were completed after a safety refusal by whatever the harness fell back to. The score is the harness's; the model credit is shared.

The benchmark owner's harness ties for first. Cognition built FrontierCode. Cognition's Devin Fusion CLI shares the top row. AA discloses it; so do we, for the second index in a row.

Do not read v2.0 next to v2.1. Every configuration fell on the new FrontierCode, by an average of five Index points. Our October post carries v2.0 numbers and now says so.


Verified vs Unconfirmed: The Scorecard

ClaimVerdict
v2.1 announced Nov 14, 00:04 UTC; components and weightsVerified (AA thread; coding agents page text)
Leaders at 58; next rows at 56 and 55Verified (AA thread)
Fable 5.2 effort ladder and costs; Sol at $1.51; 30x rangeVerified (AA thread)
Time-limit zeros; 3x tokens at maxVerified (AA thread)
Component scoresVerified (AA thread)
Fallback percentagesVerified (AA thread)
FrontierCode-AA 2.0 drops; Terminal-Bench 4.1 movesVerified (AA thread)
Per-configuration page tableNot rechecked (JavaScript-rendered)

Who Should Use It

Use Fable 5.2 at high in Claude Code if you want the top score at the lowest price on this board, $7.51 a task, and leave max alone; it scores lower and costs six times more.

Skip the fallback-free reading of any Anthropic row; ask your own harness what it does on a refusal, because AA's did something, and skip comparing any number here to the October board.

The frontier is elsewhere: on the cheapest end of the Pareto line sits Codex with GPT-6.1 Sol at 52 for $1.51, and on the Terminal-Bench component alone the leader is Kimi Code CLI with Kimi K3.1 Pro. Our model comparison has the current standings side by side.


Sources


*Not sure which effort level your coding agent should run at? Take the free 2-minute quiz and get matched. Then read the October board these numbers replace: Coding Agent Index v2.0.*

Which model should you be using?

Three minutes, twelve questions, one defensible answer.

Take the quiz →