PowerPoint Benchmark · Results summary
Which AI makes a useful editable PowerPoint deck?
Data updated · local-v38 · Protocol 0.38.0
Claude Opus 5.5 / XHigh leads the current public configuration ranking with a pairwise win score of 97.9 / 100. Its 95% interval is 95.7–99.5.
This is one corporate-design task: making an editable eight-slide deck from supplied sources, a PowerPoint master and a slide library. It does not establish a best model for every presentation task.
Current results
The release lists 66 public model and reasoning configurations. The full rating pool contains 289 decks and 987 scheduled pairs, including retired models omitted from the public listing. The summary and full leaderboard use the same release data.
| Configuration | Win score / 100 (95% interval) | Decks | Cost / deck | Minutes / deck |
|---|---|---|---|---|
| Claude Opus 5.5 / XHigh | 97.9 (95.7–99.5) | 7 | $6.54 (n=7) | 22.4 (n=7) |
| Claude Opus 5.5 / Max | 97.4 (95.6–99) | 7 | $12.86 (n=7) | 40.0 (n=7) |
| Claude Sonnet 5.5 / Max | 96.9 (94.5–98.9) | 5 | $21.39 (n=5) | 72.9 (n=5) |
| GPT-6 Astra / XHigh | 94.3 (91.2–97.6) | 8 | $3.34 (n=8) | 10.8 (n=8) |
| GPT-6 Astra / Ultra | 94.2 (78.7–98.8) | 1 | $2.58 (n=1) | 8.2 (n=1) |
Quality versus cost
The highest score, lowest cost and shortest generation time answer different questions. Among the published configurations with operating measurements, GPT-6 Luna / Low has the lowest mean cost ($0.015, n=3) and a win score of 6.9. Claude Opus 5.5 / Low has the shortest mean generation time (1.4 minutes, n=3) and a win score of 6.7.
These are measurements of the tested runs, not current vendor price quotes. Cost and time stay attached to the same reasoning configuration as its quality score. See the cost-versus-quality map and complete leaderboard before choosing a tradeoff.
What the score can tell you
A win score is the estimated chance of beating a typical rated configuration, on a 0-100 scale. It is not a percentage of task requirements passed. Blind comparisons are fitted with a Bradley–Terry model; the full comparison also shows Elo and uncertainty.
Small samples, overlapping intervals, judging errors and the task setup limit the conclusion. A new deck pool can change relative scores. Inspect the rendered decks and method and limitations, including the distinction between visual judgments and editability checks.
Citing and reusing this release
Cite Andrew McGehee, PowerPoint Benchmark, local-v38, data updated October 25, 2026, with a link to this page. The JSON snapshot includes version, scope, intervals, sample counts and evidence links. It refreshes with the publication; save the file if you need the exact snapshot.
The existing content terms allow quoting, linking and sharing with attribution; ask about substantial reuse. This export contains published aggregate results. A complete licensed reproduction pack and historical archive still need a separate release review.