⚖️ Judge verdict
The AI Cold War's Newest Weapon: Stealing the Teacher's Answer Key · Judge: anthropic:claude-opus-4-8 · 2026-09-04T20:55:58
· ✅ saved (e4bd4c) · 3 models
🔬 View outputs side-by-side →
all verdicts →
🏆 Quality winner: openrouter:deepseek/deepseek-v4-pro
Run 2eb734 wins on analogy quality (the Tu-4/B-29 Cold War reverse-engineering analogy precisely mirrors distillation's mechanism—copying that's cheaper but doesn't fully match the original), the richest and most self-contained rewritten passage, and a complete 10-question quiz whose questions are all answerable from its own passage. Run 0313fc is very close but its 'surreptitious' vocab question references a word not in its rewritten passage. Run a0132a is solid but weaker with only 3 quiz questions.
💰 Best value: openrouter:deepseek/deepseek-v4-pro
— quality 9/10 at 0.62¢ → value 8.7/10
Value = 0.7·quality + 0.3·cheapness. Quality stays dominant, so a cheap-but-weak output can't win on price alone.
Quality ranking
| # | Model | Facts | Analogy | Age-fit | Clarity | Quiz | Glossary | Passage | Overall | Cost | Value | Notes |
|---|
| 🥇 |
openrouter:deepseek/deepseek-v4-pro
view ↗ |
9 | 9 | 9 | 9 | 9 | 9 | 9 |
9 |
0.62¢ |
8.7 |
Excellent mechanistic analogies (Tu-4/B-29 reverse-engineering, filming opponent's practice, forking code) that share the actual distillation mechanism. Ten well-constructed SAT-style questions including evidence-pairing, all answerable from the rich passage. Glossary is precise and comprehensive; dates and names all accurate. |
| 🥈 |
anthropic:claude-haiku-4-5-20251001
view ↗ |
9 | 8 | 9 | 9 | 8 | 9 | 8 |
8 |
2.57¢ |
6.8 |
Strong faithful reporting with good structure and ten quiz questions. Q4 asks about 'surreptitious' which appears in source but not in the rewritten passage, slightly unfair for a passage-based quiz. Analogies (arms race, doping scandal) are good; passage is accurate and engaging. |
| 🥉 |
openrouter:baidu/ernie-4.5-vl-424b-a47b
view ↗ |
8 | 7 | 8 | 8 | 6 | 8 | 8 |
7 |
0.43¢ |
7.9 |
Solid factual grounding and clean structure. Only 3 quiz questions (vs 10 for others), and Q2's 'distil' answer choosing 'replicate' over 'condense' is debatable. Analogies (recipe, plagiarism) are decent but generic; passage is accurate and engaging. |
Scores 0–10 · 🟩 ≥8 · 🟨 4–7 · 🟥 <4 · sorted by Overall. Hover a column header for its full rubric name.
💰 Cost-adjusted ranking (value for money)
| # | Model | Quality | Cost | Value |
|---|
| 🥇 |
openrouter:deepseek/deepseek-v4-pro |
9 |
0.62¢ |
8.7 |
| 🥈 |
openrouter:baidu/ernie-4.5-vl-424b-a47b |
7 |
0.43¢ |
7.9 |
| 🥉 |
anthropic:claude-haiku-4-5-20251001 |
8 |
2.57¢ |
6.8 |
🪙 Best under 3¢ (cost < $0.03), ranked by quality
| # | Model | Quality | Cost | Value |
|---|
| 1 |
openrouter:deepseek/deepseek-v4-pro |
9 |
0.62¢ |
8.7 |
| 2 |
anthropic:claude-haiku-4-5-20251001 |
8 |
2.57¢ |
6.8 |
| 3 |
openrouter:baidu/ernie-4.5-vl-424b-a47b |
7 |
0.43¢ |
7.9 |
Strong quality without paying flagship prices — the cheap-and-good picks.