⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 5 min read

Fable 5.2 vs Opus 5.5: Every Anthropic Benchmark, Accuracy per Dollar and the GoldieBench Record

Anthropic's published results for Fable 5.2 and Opus 5.5, read against cost per task and GoldieBench's own data.

A rising gold curve over a benchmark grid with two glowing markers over dark navy

Fable 5.2 vs Opus 5.5 is a clean sweep on Anthropic's published table, with Fable 5.2 ahead on every row, but GoldieBench has only benched Opus 5.5 so far, at an average of 7.57 across 50 one-shot builds.

Book a free SEO strategy session →

This page separates Anthropic's numbers from our own and explains which effort settings decide the result.

To see how we test tools for SEO work, book a free SEO strategy session.

Vendor figures are from Anthropic's Fable 5.2 launch post and model page.

Anthropic's scoreboard

BenchmarkFable 5.2Opus 5.5Gap
FrontierCode 2.049.1%45.5%+3.6
FrontierSWE v269.9%62.3%+7.6
AutomationBench51.4%42.5%+8.9
OSWorld 2.1 (offline, partial)89.3%87.2%+2.1
Chartography (no tools)73.4%64.4%+9.0
BenchCAD 2.1 (Max effort)83.565.3+18.2

Scores show each model at its highest available effort with safeguards on, unless Anthropic noted otherwise.

The Fable 5.2 FrontierCode figure is at Xhigh effort; at High it scored 45.3%.

Accuracy against cost

Anthropic plotted accuracy against cost per task for several tests, and those charts change how to read the table.

On FrontierCode, Fable 5.2 at High effort sits level with Opus 5.5 at its best for roughly half the cost.

On WANDR, Fable 5.2 at Medium matches Opus 5.5's top score for about 40% of the cost.

On Chartography, Fable 5.2 at Medium scores 66.4%, above Opus 5.5 at Max effort, for under a tenth of the cost per task.

On these tests, the stronger model is also the cheaper way to reach a given score.

What GoldieBench shows today

On the GoldieBench leaderboard, Claude Opus 5.5 averages 7.57 out of 10 over 50 one-shot builds and sits at #13 overall, strongest in Sims.

The earlier Claude Fable 5 averages 8.10 over 47 tasks at #5, ahead of Opus 5.5.

Fable 5.2 has no runs yet, so the Fable 5 record is the closest bench evidence of how the Fable line handles one-shot builds.

Book a free SEO strategy session →

Settings behind a fair run

  • Effort: Fable 5.2 defaults to high, Opus 5.5 to medium; a fair run fixes the same level on both.
  • Thinking: adaptive and always on for both.
  • Output: 128K max on both, far more than a single HTML file needs.
  • Sampling: Fable 5.2 does not accept temperature, top_p or top_k.

Model choice by task

TaskLean towardsBasis
One-shot web buildsFable lineFable 5 8.10 vs Opus 5.5 7.57 on GoldieBench
Hard coding at volumeFable 5.2 at HighOpus-level quality at about half cost per task
Routine editsOpus 5.5 at mediumLower per-token price
Charts and visualsFable 5.2Largest Chartography gap

Price levers behind the cost charts

Anthropic's pricing page lists Fable 5.2 at double Opus 5.5's per-token rate, with both halved on batch processing.

Cached input is billed at 2.5% of base on Fable 5.2 and 5% on Opus 5.5.

Fable 5.2 is also cheaper per token than Fable 5.1, which matters for anyone comparing against our Fable 5 runs.

None of that changes the scores, but it explains why the accuracy-per-dollar charts favour Fable 5.2 at lower effort.

Spec sheet for builders

SpecFable 5.2Opus 5.5
Released9 Nov 202622 Sep 2026
Context window1M tokens1M tokens
Max output128K tokens128K tokens
API default efforthighmedium
Claude Code default effortmediummedium
Fast modeUp to 2.5x speed at premiumNot covered here

Specs are from Anthropic's models overview and the Opus 5.5 page.

How we will run Fable 5.2

  • Same fixed prompts as every other model on the board, one shot each.
  • Effort pinned and logged for each run, since the API default is high and Claude Code's is medium.
  • Build time recorded next to each score, with fast mode off for comparability.
  • Every result published with a playable demo, as with the Opus 5.5 runs.

Book a free SEO strategy session →

That way Fable 5.2 sits on the same footing as the 50 Opus 5.5 builds already on the board.

Small print in the vendor numbers

Anthropic's table reports results with safeguards on, and each model at its highest available effort unless noted.

Fable 5.2 blocks most cybersecurity tasks, which affects any security-style row, while the unrestricted Mythos 5.2 is only for verified organisations.

Those details are why we prefer our own fixed-prompt runs for builder decisions.

Where the Anthropic line sits on our board

Among Anthropic models already benched, Claude Opus 5 leads at 8.27 and #2 overall, followed by Fable 5 at 8.10 and #5, with Opus 5.5 at 7.57 and #13.

That ordering is a reminder that a newer model is not automatically a better one-shot builder.

Fable 5.2 will have to beat both earlier Anthropic entries on the same prompts to top the Anthropic group.

What vendor gaps predict for builds

The largest gaps in Anthropic's Fable 5.2 table, hardware design and chart reading, are visual and spatial tasks.

One-shot builds of games and interactive pages reward the same skills, so we expect Fable 5.2's strengths to show on our board.

The smaller computer-use gap suggests the two will play-test their own builds similarly; Opus 5.5's model page highlights computer use as a core strength.

Verdict for now

Anthropic's numbers put Fable 5.2 ahead everywhere, and our Fable 5 record suggests the line builds well one-shot; we will add Fable 5.2 runs to confirm.

For agent setups that use both, the AI Profit Boardroom has the trainings and the Agent OS.

Book a free SEO strategy session →

Read the cost charts alongside the scores and Fable 5.2 vs Opus 5.5 looks very different; for data-led SEO, book a free SEO strategy session.

FAQ

Has GoldieBench tested Fable 5.2?

Not yet. Opus 5.5 is benched at a 7.57 average and the earlier Fable 5 at 8.10.

Does Fable 5.2 beat Opus 5.5 on every benchmark?

On Anthropic's published table, yes, with the smallest gap on OSWorld 2.1 computer use.

Is Fable 5.2 cheaper per task?

On several of Anthropic's charts, yes: at High effort it matched Opus 5.5's best FrontierCode score for roughly half the cost per task.

Which effort should a fair test use?

The same level on both, because the defaults differ: high on Fable 5.2 and medium on Opus 5.5.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week