---
type: "post"
id: "claude-fable-5-2-takes-1-on-the-67e6f8"
slug: "claude-fable-5-2-takes-1-on-the-67e6f8"
title: "Claude Fable 5.2 takes #1 on the Vals Index at 71.31% accuracy"
description: "Vals AI reports that Claude Fable 5.2 ranked first among 36 models on its Vals Index v2.2, scoring 71.31% at a listed test cost of $17.94."
retrieval_nugget: "Vals AI reports that Claude Fable 5.2 ranked first among 36 models on its Vals Index v2.2, scoring 71.31% at a listed test cost of $17.94. The result is a useful, task-specific snapshot for builders comparing models, but its uneven benchmark results argue against treating one aggregate rank as a general capability verdict."
published_at: "2026-11-10"
updated_at: "2026-11-10"
record_date: "2026-11-10"
date_kind: "published_at"
topics: ["claude","anthropic","gemini"]
entities: ["Vals AI","Claude Fable 5.2","Anthropic","Claude Sonnet 5.5","Gemini 4 Carbon","Claude Opus 5.5","GPT-6 Astra"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "Claude Fable 5.2 ranks first on the Vals Index with 71.31% accuracy, a separate benchmark result from its Epoch Capabilities Index score."
prior_refs: ["https://newruntime.com/posts/claude-fable-5-2-scored-174-on-the-2701ba/"]
source_url: "https://www.vals.ai/models/anthropic_claude-fable-5-2"
source_urls: ["https://www.vals.ai/models/anthropic_claude-fable-5-2"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:claude-fable-5-2-takes-1-on-the-67e6f8"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/claude-fable-5-2-takes-1-on-the-67e6f8/","markdown":"https://newruntime.com/posts/claude-fable-5-2-takes-1-on-the-67e6f8.md","json":"https://newruntime.com/posts/claude-fable-5-2-takes-1-on-the-67e6f8.json"}
markdown_url: "https://newruntime.com/posts/claude-fable-5-2-takes-1-on-the-67e6f8.md"
---

# Claude Fable 5.2 takes #1 on the Vals Index at 71.31% accuracy

## Retrieval answer

Vals AI reports that Claude Fable 5.2 ranked first among 36 models on its Vals Index v2.2, scoring 71.31% at a listed test cost of $17.94. The result is a useful, task-specific snapshot for builders comparing models, but its uneven benchmark results argue against treating one aggregate rank as a general capability verdict.

## Summary
Vals AI reports that Claude Fable 5.2 ranked first among 36 models on its Vals Index v2.2, scoring 71.31% at a listed test cost of $17.94. The result is a useful, task-specific snapshot for builders comparing models, but its uneven benchmark results argue against treating one aggregate rank as a general capability verdict.

## What happened
Vals AI says it evaluated Anthropic’s Claude Fable 5.2 on Vals Index v2.2, its component benchmarks, and the Vals Science Index. The page lists a release date of November 9, 2026, a 1-million-token context window, 128,000 maximum output tokens, private weights, and text, image, and file input support; video input is not supported. Vals lists token prices of $8 per million input tokens and $40 per million output tokens. Its Vals Index test cost is $17.94, with a reported latency of 43 minutes and 42 seconds.

On the overall Vals Index, the model scored 71.31% with a stated margin of ±0.96, ranking first of 36. Vals places it ahead of Claude Sonnet 5.5 at 70.85% and Gemini 4 Carbon at 70.45%. The page also reports a 37.88% score on the Vals RSI Index, first of 27, and a 63.74% score on the Vals Science Index, second of 24 behind GPT-6 Astra at 65.09%. The overall ranking is therefore not a claim that Fable 5.2 leads every category.

The strongest listed results include first place on Vibe Code Bench v1.1, with 93.77% ±1.15; Finance Agent v2.1, at 69.47% ±1.98; EMB v1.1, at 77.38% ±2.19; and BioMysteryBench, at 83.70% ±0.74. It also tied Claude Opus 5.5 for first on Terminal-Bench 4.0 with 65.15% ±2.31, and ranked first of 66 on ProgramBench with 19.00%. Vals reports second place on SWE-bench Multimodal at 62.29%, behind Opus 5.5 at 70.21%.

The weaker results matter just as much when choosing a model. Fable 5.2 scored 5.00% on HLAB-Vals-1.1, tied for 29th of 37, and 32.09% on Tax Agent Bench, ranking 27th of 84. On Legal Research Bench it placed 19th of 93 at 46.15%; on the Vals Legal Index it was 20th of 37 at 69.17%. Vals says it tested at High effort with default temperature, top-p, and top-k settings, while warning that some benchmarks may use different providers and parameters.

## What the archive adds
New Runtime’s earlier coverage on [Claude Fable 5.2’s Epoch Capabilities Index score](https://newruntime.com/posts/claude-fable-5-2-scored-174-on-the-2701ba/) on November 10, 2026, recorded a different benchmark result for the same model. Earlier coverage of [Claude Fable 5.1 and GPT 6.1 Sol on a live agent leaderboard](https://newruntime.com/posts/gpt-6-1-sol-max-and-claude-fable-b17123/) and [Claude Opus 5.5 on Vals AI’s SWE-bench Multimodal](https://newruntime.com/posts/claude-opus-5-5-tops-vals-ai-s-3c4a0e/) offers relevant comparison context. This is a new evaluation fact about a known release, not an independent confirmation of broad model superiority.

## Trend signal
This result supports the pattern that [agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/): Vals presents task performance alongside cost per test, rather than only token price. The evidence is still limited to Vals’s benchmark setup. For teams, the useful comparison is whether the task success is worth the measured cost and time in a representative workflow, not whether a model tops one aggregate leaderboard.

## Why it matters
A high rank can help shortlist a model, but it should not settle a production choice. The scores vary sharply by task: Fable 5.2 is strong on Vibe Code Bench and Finance Agent, while its listed Tax Agent Bench and HLAB-Vals-1.1 positions are weak. Teams building coding or finance workflows may see a reason to run their own evaluation; teams working in legal, tax, or other domains should not infer likely performance from the overall 71.31% alone.

The cost figure is also worth reading carefully. Vals’s $17.94 is the cost of its Vals Index test, not a universal per-task price. Token charges depend on how much a workflow consumes, and the page’s latency reflects benchmark execution rather than necessarily the response time of an individual request. A practical comparison should include the same tasks, tools, prompts, effort settings, and success criteria across candidate models. The ranking is a starting signal for that test, not a substitute for it.

## Limits and uncertainty
Vals AI’s scores describe its own benchmarks and evaluation conditions; they do not establish performance across every application, prompt, tool configuration, or production workload. The page notes that some benchmarks may use different providers and parameters, so scores across categories are not necessarily directly comparable. The listed uncertainty margins are benchmark-specific and do not quantify how well the results transfer to a particular team’s tasks. Vals does not establish from these numbers whether Fable 5.2 will be more reliable or cost-effective in a given deployed workflow. The results have not been independently reproduced here.

## Skill to build now
Turn the benchmark into a local model-selection test: choose a small set of real tasks that represent your workflow, define what counts as a correct and usable completion, and record cost and elapsed time alongside success. Compare Fable 5.2 with your current model under the same tool access and effort settings. Include failure-prone cases, not only tasks resembling the strongest published benchmarks; keep the published rank as context rather than your acceptance criterion.

## Reader takeaway
Claude Fable 5.2 leads Vals AI’s overall index and several specific tests, but the same page shows pronounced weaknesses elsewhere. Use the result to decide what to evaluate next, not what to deploy without testing: the relevant question is whether the model improves your own task outcomes at an acceptable cost and latency.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)