ModelBrief · Claude Fable 5.2 · Topic Page
Last updated Nov 20, 2026

Claude Fable 5.2 vs Opus 5.5 and Gemini 4 Carbon: Twice the Price — Where the Money Goes

ModelBrief · Published 2026-11-15 · Fact-check 2026-11-15 · Reference: The Decoder

"Fable 5.2 vs Opus 5.5" is the strongest search cluster of launch week, carrying three sub-intents: token usage, benchmark comparison, and coding. The one-line verdict first: Fable is stronger, but its strength shows on the hardest tasks; on everyday work, Opus 5.5 is cheaper, faster, and outright wins two high-frequency scenarios. Carbon wins on a different axis: business-process automation.

The one-glance table

DimensionFable 5.2Opus 5.5Gemini 4 Carbon
AA Intelligence Index65 (Max)59 (Max)60 (High)
Price (input / output)$8 / $40$4 / $20$2 / $10 (intro price)
Cost per index taskMax $26.54; High $9.08About $10.2About $5.48
FrontierCode 2.049.1% (Xhigh)45.5%41.1% (Medium tier)
AutomationBench51.4%42.5%53.9%
Terminal-Bench 4.0 (AA run)59.6%59.6%65.2%
AA-Briefcase Elo / rubric1785 / 66.1%1800 / 61.2%1611 / 61.8%
Remote Labor Index41.3%21.3%—

Coding: Fable wins the benchmarks, Opus wins the daily work

On vendor-reported coding benchmarks, Fable leads across the board: FrontierCode 2.0 49.1% (Xhigh) vs Opus 45.5%, FrontierSWE v2 69.9% vs 62.3%, SWE-bench Pro 93.0% (Max, system card table 8.1.A) vs 89.9%, CursorBench 4.0 60.0% (Max) / 56.9% (High). The system card says Fable wins 14 of 16 headline results. But independent tests paint a more mixed picture: in FinOps' six-task coding run, Opus 5.5 and GPT-6.1 Sol passed 6/6 while Fable 5.2 passed 5/6 — a small sample, not a ranking, but a reminder that Fable's lead doesn't cash out in every coding scenario.

The decisive numbers sit in the FrontierCode tier sweep: Fable High at 45.3% / $6.72 nearly ties Opus Max at 45.5% / $12.82, at about half the cost per task. The same coding score, for half the money, on Fable High rather than Opus Max — but Fable Max (45.7% / $44.70) scores below its own Xhigh (49.1% / $14.19). The coding verdict: Fable High is the coding sweet spot; Max is not.

Usage in Claude Code splits by tier too: it requires 2.1.321+, /effort switches tiers, and the default is medium. The community consensus is split routing — "Fable 5.2 plans, Opus 5.5 executes" (Matt Shumer, about 2,800 likes); another widely shared line is "Opus 5.5 is not suddenly obsolete… Fable 5.2 is just a better boss" (@rohit3a — Fable Medium orchestrates, Opus executes). Both point to the same pattern: the expensive model plans, the cheap one executes.

Token usage and per-task cost: where the money goes

Per token, Fable costs about twice Opus ($8/$40 vs $4/$20). Per task, the gap is wider: AA's cost per index task at Max is $26.54, about 2.6× Opus — because Fable is not only dearer per token but also more verbose and makes more calls. On VIALS-AA, Fable averaged 37.4 tool calls per task, the most of any Anthropic model; in ComputingForGeeks' test, medium-tier Fable wrote twice the tokens of 5.1. Casagbic's SQL test is the cautionary case: across 1,280 queries, Fable scored 320/320 at $0.029 per query while Haiku 5.5 scored 319/320 at about 54× lower cost — on accuracy-ceiling tasks, Fable's premium is nearly pure waste.

Time splits the same way: AA time per task is about 45.3 minutes at Max, 27.7 at Xhigh, 16.1 at High (Nov 15, 2026 reading, drifting), with time to first token around 268 seconds at Max. On everyday tasks, Opus is not only cheaper — it gets to the answer sooner.

Carbon: the AutomationBench upset, at a lower price

Gemini 4 Carbon (AA 60, High tier) is Fable's nearest rival on the board and the only model to beat it head-on in the official table: AutomationBench 53.9% vs 51.4%. On AA's independent Terminal-Bench 4.0 run, Carbon's 65.2% also beats Fable's 59.6%. On price, Carbon is $2/$10 on an introductory rate, about $5.48 per index task — under a quarter of Fable Max. Carbon's weak spots are FrontierCode (41.1%, run at Medium, conditions not fully equal) and AA-Briefcase Elo (1611). Business-process automation and terminal tasks point to Carbon; long-horizon agentic coding and scientific reasoning point to Fable High.

In brief: GPT-6 Astra and GLM 5.2

GPT-6 Astra (AA 57) costs $6.24 per index task, under a quarter of Fable Max; on FrontierSWE, Fable leads 69.9% to 65.5% — a real but modest gap. The Fable vs GLM 5.2 comparison appears in both English and Chinese autocomplete ("claude fable vs glm 5.2"), but as of Nov 15 no dedicated comparison article exists in public channels — this site won't invent scores, and will add it once a checkable comparison appears.

Decision tree: when to use which

Everyday coding and knowledge work → Opus 5.5 (Anthropic's official advice; cheaper, faster, higher AA-Briefcase Elo). The hardest long-horizon agent, science and research tasks → Fable 5.2 High (62 points, $9.08/task; ties Opus Max for about half the cost). Tasks where xhigh beats max, such as FrontierCode and ARC-AGI-2 → Fable 5.2 Xhigh (49.1% / 93.8%). Business-process automation → Carbon (AutomationBench 53.9%, cheaper). Only tasks worth $26.54 where High/Xhigh clearly fall short → Fable 5.2 Max. Accuracy-ceiling tasks (SQL queries, format conversion) → a light model (Haiku-class); Fable's premium is pure waste there.

ModelBrief's take: Fable 5.2 vs Opus 5.5 isn't a coronation and a dethroning — it's a division of labour: Fable is the stronger boss, Opus the more economical executor. Twice the price buys the ceiling on the hardest tasks, not a blanket win on everyday work. And Carbon is the reminder that the 65-point headline doesn't include business-process automation.

Reference: The Decoder — Claude Fable 5.2 and Mythos 5.2 cement Anthropic's AI lead just before its IPO (rewritten and organized by ModelBrief; facts per official sources and Artificial Analysis)