Appearance
Benchmarks
Recall on a real chat
One real Korean conversation: 285,000 characters over 170 messages. 40 questions written from it: 30 about a single scene and 10 that need two scenes far apart. Claude Opus 5.5 answered every row and the grading was the same for all three.
| Method | Input tokens | Single scene (30) | Two scenes (10) | Total |
|---|---|---|---|---|
| Full history as text | 321k / turn | 30 | 8 | 38 |
| fiche, 16 px pages | 84k / turn | 28 | 8 | 36 |
| RAG: voyage-4-large + rerank-3, top 20 | 11k / question | 30 | 7 | 37 |
- When fiche missed, it said it didn't know instead of inventing an answer.
- RAG missed one of the two scenes on 3 of the 10 two-scene questions, and giving it up to 180 passages didn't fix that.
- Same test on Gemini with 8,000-character PDF pages: 38.5 / 40 on about 43,000 input tokens.
Density and accuracy
Rare words hidden in a page, 30 per setting. Denser pages save more and misread more.
| Model | Setting | Tokens saved | Exact-word recall |
|---|---|---|---|
| Gemini | 6,000 chars / page | ~71% | 100% |
| Gemini | 8,000 chars / page | ~78% | 98.5% |
| Gemini | 10,000 chars / page | ~83% | 95% |
| Claude | 18 px | ~67% | 96.7% |
| Claude | 16 px | ~75% | 93.3% |
| Claude | 14 px | ~80% | 63.3% |
TIP
Measured November 2026 on small samples. Savings are against each model's own text tokenizer.