New page, November 11, 2026
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Every figure here was re-read from the launch post and the Artificial Analysis leaderboard on November 11, 2026, and each table states its source and date.
Fable 5.2 vs Opus 5.5: the full comparisonVendor + independent data · verified November 11, 2026
Claude Fable 5.2 benchmarks
Two tables, two sources. The first is Anthropic’s own launch table from November 9, 2026 — five models on one harness. The second is Artificial Analysis’s independent index, read on November 11, 2026. Neither settles your workload; the caveats section says where each one stops.
The short answer
On Anthropic’s own table, Fable 5.2 beats Opus 5.5 on 11 of 12 rows: +7.6 points on FrontierSWE v2 (69.9% vs 62.3%), +20.0 on the Remote Labor Index (41.3% vs 21.3%), and +16.7 on FrontierCode 2.0 against its own predecessor (49.1% vs 32.4%). It does not lead everything — Opus 5.5 keeps AA-Briefcase Elo (1800 vs 1785) and Gemini 4 Carbon wins AutomationBench (53.9% vs 51.4%). The first independent numbers, from Artificial Analysis on November 11, 2026, put Fable 5.2 top of its index at max effort (65 against Opus 5.5’s 59 and Gemini 4 Carbon’s 60) for $26.54 per index task against $10.04.
Anthropic’s launch table, verbatim
Twelve rows from Anthropic’s November 9, 2026 announcement — five models on the same harness, each at its highest available effort with safeguards on unless the cell carries an effort label. Vendor-reported: Anthropic built the harness, chose the tasks and ran the models.
| Claude Fable 5.2 | Claude Fable 5.1 | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon | |
|---|---|---|---|---|---|
| FrontierCode 2.0 (agentic coding) | 49.1% (xhigh) | 32.4% (xhigh) | 45.5% | 42.8% | 41.1% (medium) |
| FrontierSWE v2 (agentic coding) | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench (business workflows) | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 (agentic research) | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% |
| WANDR, relaxed score (agentic search) | 82.1% | 68.7% | 72.3% | — | — |
| OSWorld 2.1, offline partial set (computer use) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim (robotics) | 64.9% | 27.6% | 39.4% | 29.8% | — |
| AA-Briefcase v1.1 (knowledge work, rubric) | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| AA-Briefcase v1.1 (Elo) | 1785 | 1669 | 1800 | 1569 | 1611 |
| Remote Labor Index (knowledge work) | 41.3% (high) | 17.9% | 21.3% | 20.8% | — |
| Humanity’s Last Exam, with tools | 74.2% | 65.6% | 67.7% | 57.2% | — |
| Chartography, no tools (chart recognition) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Source: Anthropic, “Introducing Claude Fable 5.2 and Claude Mythos 5.2”, November 9, 2026. Three rows from that table are deliberately absent — CyScenarioBench, CVE Rediscovery and de novo binder design — because they score Mythos 5.2 against Opus 5.5 “without safeguards”, so they are not Fable 5.2 numbers. Effort labels are Anthropic’s own: Fable 5.2 and Fable 5.1 on FrontierCode 2.0 at xhigh, Fable 5.2 on the Remote Labor Index at high, Gemini 4 Carbon on FrontierCode 2.0 at medium. A dash means the vendor published no figure.
Artificial Analysis, read November 11, 2026
The first independent evaluation of Fable 5.2 that was public two days after launch. Intelligence Index v5.0.1 bundles nine evaluations — AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, the CritPt v2 hard subset and Terminal-Bench-Science 0.1. Scores are the integers Artificial Analysis displays.
| Intelligence Index v5.0.1 | Cost per index task | AA-Briefcase Elo | |
|---|---|---|---|
| Claude Fable 5.2 (max effort) | 65 | $26.54 | 1785 |
| Claude Fable 5.2 (xhigh) | 62 | $14.83 | 1738 |
| Claude Fable 5.2 (high — API default) | 62 | $9.08 | 1677 |
| Claude Fable 5.2 (medium) | 56 | $3.80 | 1622 |
| Claude Fable 5.2 (low) | 54 | $3.10 | 1563 |
| Claude Opus 5.5 (max) | 59 | $10.04 | 1801 |
| Gemini 4 Carbon (high) | 60 | $7.76 | 1611 |
| GPT-6 Astra (max) | 57 | $6.24 | 1569 |
| Claude Sonnet 5.5 (max) | 55 | $10.02 | 1816 |
Source: Artificial Analysis model leaderboard (artificialanalysis.ai/leaderboards/models), read November 11, 2026. Cost per index task is AA’s weighted average price of running one Intelligence Index task at list rates — it measures token spend on AA’s harness, not your bill. Elo is AA-Briefcase v1.1, the same metric Anthropic quotes. AA publishes decimals and this page rounds to whole numbers, so Opus 5.5 reads 1801 here against the 1800 Anthropic printed in the table above — two sources, one number to within rounding.
What the numbers actually say
Three readings that survive both tables.
The lead over Opus 5.5 is real but uneven
The gaps concentrate on long-horizon work: +20.0 points on the Remote Labor Index, +15.9 on Terminal-Bench-Science, +7.6 on FrontierSWE v2. They shrink on computer use (+2.1 on OSWorld 2.1) and reverse on knowledge-work Elo, where Opus 5.5 leads by 15. Independently, AA’s composite gap at max effort is 6 points (65 vs 59) — the same direction as Anthropic’s table, measured on a different harness. One caveat the table shows plainly: Sonnet 5.5 tops the AA-Briefcase Elo column at 1816 while sitting 10 points lower on the composite index, so a single rubric Elo is not a ranking.
Effort moves the score more than the model does
Fable 5.2 spans 54 to 65 on AA’s index across effort levels — an 11-point range — while the API defaults to high. At high it scores 62 for $9.08 per index task; at max, 65 for $26.54. Opus 5.5 at max scores 59 for $10.04. If the budget is fixed, the effort setting is the variable worth sweeping before you switch models.
Against Fable 5.1 it is a price cut as well as a jump
Fable 5.2 scores 49.1% on FrontierCode 2.0 against Fable 5.1’s 32.4% — +16.7 points — and 69.9% vs 56.3% on FrontierSWE v2, while list price fell 20% to $8/$40 per million tokens. Anthropic also reports output up to 30% faster and about 55% fewer safeguard blocks than 5.1. The two generations are not comparable on AA’s index: Fable 5.1 is marked deprecated and carries no current index score.
Five things these tables do not tell you
Read this before quoting any number on the page.
Table one is vendor-reported
Anthropic designed the harnesses, picked the tasks and ran the models. That does not make the numbers wrong, but a vendor table answers “what can this model do on our evaluation”, not “will it beat the alternative on my work”. No row in the first table has been independently reproduced cell by cell.
Only compare within a table
FrontierCode 2.0 appears in both tables and the numbers differ: 49.1% in Anthropic’s, 48.8% on AA’s FrontierCode-AA 2.0 at max effort. Same family of tasks, different scaffolding, different scoring. Likewise Anthropic’s 74.2% on Humanity’s Last Exam and AA’s HLE-Diamond 80.2% are two different tests, not a disagreement.
The loudest launch numbers are Mythos, not Fable
CyScenarioBench 76.3%, CVE Rediscovery 89.6% and de novo binder design 85.1 are Claude Mythos 5.2 — the same underlying model without the cyber and biology safeguards, sold only to vetted organizations. Fable 5.2’s safeguards block most of those tasks, so quoting them as Fable 5.2 scores would be wrong.
No arena Elo yet
As of November 11, 2026 we could not read a human-preference Elo for Fable 5.2: the LMArena leaderboard redirects to arena.ai behind a bot check, and the public SWE-bench leaderboard carried no Fable 5.2 entry. Independent coverage so far is Artificial Analysis’s index. Arena and SWE-bench figures will be added only when a page we can actually read shows them.
Cost per task is not your cost
AA’s $26.54 per index task is list-price arithmetic on AA’s token counts at max effort. Your tasks, cache hit rate, batch discount and effort setting move it by multiples — a cached, batched workload at high effort can land an order of magnitude lower.
Every figure on this page was re-read on 2026-11-11. Prices come from the model registry; benchmark scores are quoted as published and never recomputed.
Decide with the price next to the score
A benchmark lead is worth what it saves you. The comparison page has the full spec and cost picture; the calculator prices your own token volumes.