PPT App Testing: Cost Anomaly Across Model Multipliers (Noticed Along the Way)
While testing the PPT App today (single-page PPT generation), noticed some inconsistencies between token cost and the model multiplier numbers — flagging it here.
These weren’t dedicated cost-measurement tests — just observations picked up during regular PPT testing. But since each run used the same prompt and same variables (search on, same template choice or none), with only the model/multiplier changed, the data still seems worth sharing.
Gemini ×3
- 1-1: no template — 221 tokens / 10min
- 1-2: no template — 255.28 tokens / 7min
Gemini ×6
- 2-1: no template — 299 tokens / 10min
- 2-2: template — 602 tokens / 14min
- 2-2: template — 602 tokens / 14min
- 2-5: youming template — 523.455 tokens / 20min18s
Gemini ×18
- 3: no template — 2717.623 tokens / 13min
GPT ×4
- 4-1: template — 177.629 tokens / 9min
Kimi ×6.78
- 5: no template — 585.105 tokens / 20min29s
Qwen ×3.31
- 6: no template — 1108.927 tokens / 15min35s
GPT ×20
- 7: no template — 378 tokens / 6min
GPT ×28
- 8: no template — 409.358 tokens / 10min
Claude ×30
- 9: no template — 2235.757 tokens — cost is roughly in the same range as Gemini ×18 (2717.623), despite the multiplier being nearly double
The issue
Not a dedicated cost test, but since scenario (single-page PPT), prompt, and variables were all consistent — only the model/multiplier changed — cost should scale somewhat predictably with the multiplier. It doesn’t:
- Gemini 6→18 (3x multiplier increase) → actual cost jumped ~5–9x (vs. normal ×6 baseline of ~300–600 tokens)
- Claude ×30 costs about the same as Gemini ×18, despite the multiplier being nearly double
Just something noticed along the way, not a rigorous cost comparison experiment — but the numbers diverge enough that it’s worth the team taking a look at whether there’s a billing logic issue.