ProjectsPowerPoint BenchmarkPowerPoint
Protocol 0.38

By · Compact results · Methodology and prompt · Results JSON

PowerPoint sucks.

I built a repeatable benchmark to find out which AI can produce a usable, editable corporate deck.

66 AI model and reasoning configurations from OpenAI, Anthropic and other labs each built the same editable consulting deck on one corporate PowerPoint master and slide library. Claude Opus 5.5 · XHigh currently ranks first.

View the leaderboardSee how it works

01 — Results

PowerPoint Benchmark leaderboard.

An AI judge panel compares decks two at a time. A win score of 50 means even odds against a typical model. The panel agreed with human reviewers on 92% of 49 test pairs.

Cost vs. quality

Up and to the left is better.

Best valueDeepSeek V4.1 Flash 1017 · Maxwin score 89.8 · $0.25 per task
Top qualityClaude Opus 5.5 · Xhighwin score 97.9 · $6.54 per task
  • OpenAI
  • Anthropic
  • Other
  • Frontier

02 — Models

Leaderboard.

One row per model, ranked by its best tested win score. Cost and time use that same effort setting. Expand a model for every tested effort level.

Sort by
RankProviderModelBest effortCompare
01
Anthropic
Claude Opus 5.5Xhigh97.922m 22s$6.54Compare decks
02
Anthropic
Claude Sonnet 5.5Max96.972m 56s$21.39Compare decks
03
OpenAI
GPT-6 AstraXhigh94.311m 16s$3.31Compare decks
04
OpenAI
GPT-5.6 SolMax94.017m 45s$3.05Compare decks
05
DeepSeek
DeepSeek V4.1 Flash 1017Max89.826m 55s$0.25Compare decks
06
Moonshot AI
Kimi K3.1 ProHigh89.613m 16s$1.11Compare decks
07
OpenAI
GPT-6 SolMax80.111m 02s$0.96Compare decks
08
OpenAI
GPT-6.1 SolMax79.322m 43s$0.89Compare decks
09
Anthropic
Opus 5Xhigh76.627m 49s$10.39Compare decks
10
Anthropic
Claude Haiku 5.5High75.48m 37s$0.30Compare decks
11
Anthropic
Claude Opus 4.8Max75.034m 15s$11.27Compare decks
12
MiniMax
MiniMax M3.1 FlashMax70.740m 24s$0.34Compare decks
13
Anthropic
Claude Fable 5.1Max*54.822m 57s$8.65Compare decks
14
Z.AI
GLM 5.3Max47.215m 21s$0.62Compare decks
15
OpenAI
GPT-6 LunaMax44.525m 40s$0.09Compare decks
16
Xiaomi
MiMo-V2.6-FlashMax43.973m 31s$0.20Compare decks
17
DeepSeek
DeepSeek V4.1 FlashMax42.516m 45s$0.14Compare decks
18
xAI
Grok 4.6Max*42.213m 17s$2.36Compare decks
19
Z.AI
GLM 5.3 FlashMax37.027m 42s$0.24Compare decks
20
OpenAI
GPT-5.6 TerraMax34.515m 27s$1.64Compare decks
21
Qwen / Alibaba
Qwen 3.8 FlashMax31.730m 08s$0.19Compare decks
22
Moonshot AI
Kimi K3Max27.513m 38s$1.73Compare decks
23
OpenAI
GPT-5.5Xhigh18.311m 29s$2.56Compare decks
24
Google
Gemini 3.8 FlashMax13.816m 18s$1.93Compare decks
25
Meta
Muse Spark 1.3Max*13.812m 44s$1.52Compare decks
26
OpenAI
GPT-5.6 LunaXhigh4.411m 59s$0.12Compare decks
27
DeepSeek
DeepSeek V4 Pro 0813Max4.327m 15s$0.46Compare decks
28
Meta
Muse WonderMax1.711m 22s$3.03Compare decks

Ranks include all providers. Costs are mean USD per task, using provider-reported charges or API-equivalent estimates. Expand a model for sources, sample counts, and score intervals. Missing values are not ranked.

* Includes decks from runs that were stopped before the model finished; they are being replaced with fresh runs.

03 — Compare decks

Compare decks.

Model A vs Model B · Change
Claude Opus 5.5 · Xhigh97.9/ 100
Claude Opus 5.5 · Xhigh, Corporate Design, slide 1
Claude Opus 5.5 · Max97.4/ 100
Claude Opus 5.5 · Max, Corporate Design, slide 1

Slide 18

04 — How it works

How the benchmark works.

Stage 01 of 06 · freeze

Same prompt

Every model gets the same golden test case: prompt, source files, requirements, PowerPoint master, and slide library.

GateFrozen before the first run. Every model sees identical inputs.

See inputs and outputs
In
  • Prompt
  • Source files
  • Requirements
  • PowerPoint master
  • Slide library
Out
  • Golden test case
golden test case (Case 4)
1

Reading the score

What the score means.

Win score and Elo

The win score is the estimated chance of beating a typical rated configuration in a blind comparison. The Elo rating shows the same result on a chess-style scale.

Average win score
50
Average Elo
1,500
Decks · blind pairs
289 · 987
Agreement with human A/B reviews
92%
Compare any two models →

Try it: compare any two configurations

98%Claude Opus 5.5 / XHigh is expected to win 98% of blind comparisons against GPT-6.1 Sol / High.

Their 95% intervals do not overlap, so this is a real difference in this deck pool.

How the rating is computed

Two blinded AI judges, Claude Opus 5.5 and GPT-6 Sol, compare Test Case 4 decks. Their wins and ties are fitted with a Bradley–Terry model, which gives every configuration one strength rating. The 0–100 win score is that rating's chance of beating a typical configuration; Elo shows the same rating with 1500 as average. Ratings can shift when new decks join the comparison pool.

Human A/B reviews test agreement with the panel; they do not change its ratings. Models with few decks or overlapping confidence intervals may be too close to distinguish reliably.

Data caveat: 33 decks from interrupted runs

Marked results still include some of the 33 decks from runs that an earlier version of the benchmark stopped before the model could finish reviewing its work. They stay in the ratings until fresh, uninterrupted runs of the same model and effort replace them: Claude Fable 5.1 / Max (3), Grok 4.6 / Max (4), Muse Spark 1.3 / Max (4).

Judging details
PanelClaude Opus + GPT Solblind, both deck orders
ModelBradley–Terrywins and ties
DisplayWin score + Elo50 / 1500 = average

Current release: local-v38 · benchmark version 0.38.0.

PowerPoint master

Reusable layout definitions from the supplied corporate system.

0106
Swipe or use the arrows to browse layouts

Slide template

Example slides from the companion design library.

0149
Swipe or use the arrows to browse templates

Human spot-checks are called out in the published results.