---
type: "post"
id: "gamebench-2-ranks-claude-fable-5-2-first-1510a6"
slug: "gamebench-2-ranks-claude-fable-5-2-first-1510a6"
title: "GameBench 2 ranks Claude Fable 5.2 first; low-effort programs beat Fable 5.1 at lower cost"
description: "DuelLab’s November 11 GameBench 2 release ranks Claude Fable 5.2 first overall, with a 95.0 score across three reasoning settings."
retrieval_nugget: "DuelLab’s November 11 GameBench 2 release ranks Claude Fable 5.2 first overall, with a 95.0 score across three reasoning settings."
published_at: "2026-11-14"
updated_at: "2026-11-14"
record_date: "2026-11-14"
date_kind: "published_at"
topics: ["claude","anthropic"]
entities: ["DuelLab","GameBench 2","Claude Fable 5.2","Claude Fable 5.1","Claude Opus 5.5","GPT-6 Astra","Anthropic"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "GameBench 2 ranks Claude Fable 5.2 first overall at 95.0, with its low-effort programs outperforming Fable 5.1 at a lower per-program cost."
prior_refs: ["https://newruntime.com/posts/claude-fable-5-2-is-now-available-in-0cc74d/","https://newruntime.com/posts/claude-fable-5-2-takes-1-on-the-67e6f8/"]
source_url: "https://duellab.org/blog/claude-fable-5-2-debuts-in-first-place"
source_urls: ["https://duellab.org/blog/claude-fable-5-2-debuts-in-first-place"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:gamebench-2-ranks-claude-fable-5-2-first-1510a6"
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/gamebench-2-ranks-claude-fable-5-2-first-1510a6/","markdown":"https://newruntime.com/posts/gamebench-2-ranks-claude-fable-5-2-first-1510a6.md","json":"https://newruntime.com/posts/gamebench-2-ranks-claude-fable-5-2-first-1510a6.json"}
markdown_url: "https://newruntime.com/posts/gamebench-2-ranks-claude-fable-5-2-first-1510a6.md"
---

# GameBench 2 ranks Claude Fable 5.2 first; low-effort programs beat Fable 5.1 at lower cost

## Retrieval answer

DuelLab’s November 11 GameBench 2 release ranks Claude Fable 5.2 first overall, with a 95.0 score across three reasoning settings.

## Summary
DuelLab’s November 11 GameBench 2 release ranks Claude Fable 5.2 first overall, with a 95.0 score across three reasoning settings. Its results also show a practical cost trade-off: Medium scored 94.0 at an estimated $0.65 per program, while XHigh scored 99.1 at $4.58 on average. For teams comparing models, the result makes reasoning effort and cost part of the same evaluation.

## What happened
DuelLab published a dated GameBench 2 result on November 11, 2026, generated at 08:31 UTC. It says the benchmark has models write programs for eight strategy games, after which those programs play each other. The release covers 105 models, 2,237 programs, and 627,550 rated matches. DuelLab began generating Claude Fable 5.2 programs on November 9, the day Anthropic released the model.

The Overall ranking averages a model’s three reasoning settings. Fable 5.2 scores 95.0, ahead of GPT-6 Astra at 88.8, Claude Opus 5.5 at 87.6, and Claude Fable 5.1 at 87.1. These are not scores from one shared setting: the Overall figure combines Low, Medium, and XHigh. Among all 258 model-and-setting rows, Fable 5.2’s XHigh result is highest, at 99.1. That row won 4,452 of its 4,484 matches.

The setting-by-setting results sharpen the comparison. Fable 5.2 scores 91.8 at Low, 94.0 at Medium, and 99.1 at XHigh. Its Medium result comes within a point of the other three models’ XHigh results, while its Low score is 0.4 points above the best Medium score among those competitors. At XHigh, Fable 5.2 has the highest game score in seven of the eight games; its scores range from 97.6 in Go to 99.7 in Xiangqi. It won all 544 Xiangqi matches and all 554 International Draughts matches.

DuelLab reports working programs for all 24 Fable 5.2 game-and-setting pairs, with no repair requests needed. The first request was sufficient for 23 programs. One XHigh program exceeded the 128,000 output-token limit before writing code and needed another request. DuelLab estimates mean cost per program with repairs and failed attempts included. At list prices, it puts Fable 5.2 at $0.47 for Low, $0.65 for Medium, $4.58 for XHigh, and $1.90 overall. Anthropic lists Fable 5.2 at $8 per million input tokens and $40 per million output tokens, versus $10 and $50 for Fable 5.1. DuelLab says the $4.58 XHigh figure averages repeated generation runs; counting every request in full instead gives about $5.67, above Fable 5.1’s $5.15.

## What the archive adds
New Runtime’s earlier coverage reported Claude Fable 5.2 leading Cognition’s FrontierCode 2.0 benchmark at 49.1% and becoming available in Devin on November 10 ([Claude Fable 5.2 is now available in Devin](https://newruntime.com/posts/claude-fable-5-2-is-now-available-in-0cc74d/)). This is a new fact about the same model: DuelLab adds independent results across games, reasoning settings, and estimated program costs.

## Trend signal
This result supports the broader shift toward evaluating agent systems by completed-work outcomes and their costs, rather than model calls alone ([Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/)). DuelLab measures the programs’ performance in matches and estimates generation costs, but its game benchmark does not establish cost per verified task in production. The connection is therefore suggestive, not a direct measure of deployed-agent economics.

## Why now
The release follows Anthropic’s November 9 launch of Claude Fable 5.2. DuelLab says it began generating programs that day and published the benchmark result two days later. The timing gives builders an early independent comparison against several competing models, while the result remains a dated snapshot rather than a live ranking.

## Why it matters
A top score does not by itself identify the best setting for a team. In this benchmark, Medium reaches 94.0 at an estimated $0.65 per program, close to the XHigh scores of the three comparison models, while Fable 5.2’s own XHigh setting costs several times more. That makes a budget-aware evaluation more useful than treating the highest leaderboard number as a default.

The results also show why teams should track the full run, not just successful outputs. DuelLab includes repairs and failed attempts in its mean cost, and the alternative XHigh accounting that counts each request in full changes the estimate from $4.58 to $5.67. Those figures are specific to this benchmark’s programs and public list prices; they should not be used as a forecast for other coding or agent workloads. Still, the comparison is a useful model for internal tests: measure quality and cost at each reasoning setting on the tasks your team actually cares about.

## Limits and uncertainty
GameBench 2 evaluates programs in eight strategy games against a fixed reference field; its scores are not a general measure of coding quality or performance on a team’s production tasks. The 99.1 result is close to the top of the benchmark’s scale, so it leaves limited room for improvement against the current field. DuelLab’s costs are estimates based on public list prices, not provider invoices, and its XHigh average differs from the figure obtained by counting every request in full. The results are a dated release, not a live ranking, and do not establish how performance or cost will hold on other workloads.

## Skill to build now
Run a small comparison on a representative task set before assigning a high reasoning setting by default. Record successful completion rate, repair and failed-attempt costs, and total spend for each setting. Where the task permits, compare a lower-effort run with a higher-effort one; the benchmark’s Medium-versus-XHigh results show why the extra score should be weighed against its additional cost.

## Reader takeaway
Claude Fable 5.2 leads this GameBench 2 snapshot, but its highest score comes at a much higher estimated program cost than Medium. Treat reasoning effort as a quality-and-cost decision, and validate the trade-off on your own work rather than carrying a game benchmark’s ranking straight into production.

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.
- 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)