A tie at the top, on a harder test
The headline number went down, and that is the first thing to read correctly. In version 2.0, published in October, Claude Code with Opus 5.5 led at 61.7. In v2.1 the top score is 58, reported by Artificial Analysis for three model-and-harness combinations: Fable 5.2 at high and xhigh in Claude Code, Opus 5.5 at max in Claude Code, and the Opus 5.5 paired with GPT-6 Luna in Devin Fusion CLI.
That drop says more about the test than about the models. The rebuilt FrontierCode-AA 2.0 is the component that moved: all 23 configurations that were run on both versions scored lower on it, with the mean falling from 45% to 29%. Terminal-Bench 4.1, by contrast, barely moved anything — every score stayed within 3 points of its 4.0 result, while 4.1 keeps the same 66 tasks and changes the verifiers, partly to resist reward hacking. Comparing a v2.1 score with a v2.0 score is comparing two different exams.
What different settings cost
The index page also reports average pay-per-token API cost per task, and the spread is wider than the spread in scores. Among configurations scoring 48 or more, Codex with GPT-6.1 Sol (max) scores 52 and costs $1.51 per task. At the other end, Claude Code with Fable 5.2 at max effort costs $44.22 per task.
Fable's own price ladder is the sharpest detail in the release. In Claude Code it costs $2.89 per task at medium effort, $7.51 at high, $15.47 at xhigh and $44.22 at max. Its index score at high and xhigh is 58, the joint top score. At max, where it costs roughly six times the high setting, its score is 56. AA says more Fable max-effort attempts hit task time limits, so the lower score is a result under this evaluation's constraints, not proof that max effort is generally inferior. On this test, the most expensive setting of the leading model is not its best setting. Anyone buying the highest effort level by default is, on this test, paying more for a lower score.
Two cautions belong next to those prices. The costs are pay-per-token API costs; Artificial Analysis itself notes that many people reach these harnesses through subscriptions, where the per-task figure does not apply directly. And part of some headline scores was not produced by the named model alone: AA permits another model to complete an attempt after a safety refusal and includes that outcome in the configuration's score. AA reports fallback shares of about 9% of attempts for Devin Fusion and for Opus 5.5 at max, 6% for Fable 5.2 at xhigh and 5% for Sonnet 5.5, among the configurations it lists. A score with a stand-in in some attempts is still a useful score, but it is not a pure measure of one model.
### Check the summary against the source
Early secondary coverage of v2.1 described it as running FrontierCode v1.1 with 100 tasks across roughly 290 tasks in total. Artificial Analysis's own methodology page says FrontierCode-AA 2.0, 132 tasks, and 322 tasks across the three components. The secondary version carries numbers from the superseded index version. CryptoBriefing and KuCoin summaries of v2.1 are the ones that conflict with the methodology page; those figures are not used in this article.
Sources
- [1] Artificial Analysis — Coding Agent Index v2.1 methodology (components, task counts, scoring, fallback and refusal definitions, version history; read first-hand)Read source
- [2] Artificial Analysis — Coding Agent Benchmarks leaderboard (index definition, cost basis, fallback and refusal rates; read first-hand)Read source
- [3] Artificial Analysis — announcement post on X, 14 November 2026 (tie at 58, average index decline, Terra exception, per-effort costs, Fable max timeouts; read first-hand)Read source
- [4] Note on a checked discrepancy: CryptoBriefing and KuCoin summaries of v2.1 (14 November 2026) describe FrontierCode v1.1, 100 tasks and about 290 tasks total; those figures conflict with Artificial Analysis's methodology page and are not used in this article.