The same report and design system, generated by 3 models (Haiku, Sonnet, Opus) × 2 skills — my custom carousel vs. Anthropic's built-in pptx. Real model verified, time & tokens measured, every PDF inspected by eye. All 57 slides below — scroll each strip.
Note: the carousels themselves are in French (the source report's language); this interface is in English. Full write-up: Version the logic, not the deliverable.
Green rows = carousel condition; orange rows = pptx condition. Tokens = sum of subscription components (dominated by agentic cache_read; not an API-call cost). Quality scored from visual inspection.
| Model (real) | Condition | Slides | Builds | Time | Tokens Σ | QA sub. | Fonts | Quality |
|---|---|---|---|---|---|---|---|---|
| claude-haiku-4-5 | carousel | 10 | 2 | 81 s | 1.64 M | 0 | ✓ brand | 9.0 |
| claude-haiku-4-5 | pptx | 9 | 2 | 118 s | 1.56 M | 1 | ✓ brand | 7.0 |
| claude-sonnet-4-6 | carousel | 10 | 5 | 480 s | 5.47 M | 0 | ✓ brand | 9.0 |
| claude-sonnet-4-6 | pptx | 7 | 3 | 853 s | 5.37 M | 4 | ✗ substituted | 6.5 |
| claude-opus-4-8 | carousel | 11 | 2 | 353 s | 4.33 M | 0 | ✓ brand | 9.5 |
| claude-opus-4-8 | pptx | 10 | 2 | 604 s | 4.96 M | 1 | ✓ brand | 9.5 |
Carousel: quality 9.0–9.5 (spread 0.5). Pptx: 6.5–9.5 (spread 3.0). The skill compresses model variance; it is also faster and needs no QA sub-agents.
Sorted by score. Each strip scrolls horizontally (CSS scroll-snap, no JS).
Best narrative arc (myth→reality pivot); self-fixed an empty slide. Brand fonts, 353 s. · 11 slides











The most design-faithful pptx: sharp corners, 1px top-border cards, pull-lines. 604 s + 1 QA sub-agent. · 10 slides










Widest layout variety; one empty band on the steps slide. · 10 slides










SUBSTITUTED fonts (Cambria/Arial), 7 slides, 14 min + 4 QA sub-agents. · 7 slides







Brand fonts guaranteed, clean arc — in 81 s. The small model at the big ones' level. · 10 slides










Brand fonts (a nice surprise) but a centered cover + an off-spec circle. · 9 slides








