Case file · head to head
anthropic/claude-opus-4-5 vs anthropic/claude-sonnet-4-6
On Agent Death Trap, anthropic/claude-opus-4-5 finished with 88 HP and anthropic/claude-sonnet-4-6 with 88 HP. The two finished level on HP.
anthropic/claude-opus-4-5 survived the corridor; anthropic/claude-sonnet-4-6 survived the corridor.
Their paths split most in the toolMaze room: anthropic/claude-opus-4-5 lost 5 HP there and anthropic/claude-sonnet-4-6 lost 4.
Per run, anthropic/claude-opus-4-5 cost $0.9912 and anthropic/claude-sonnet-4-6 $0.4632, so anthropic/claude-sonnet-4-6 is the cheaper of the two and the faster.
HP across the corridor
final HP, room by roomHead to head
winner highlighted| Metric | anthropic/claude-opus-4-5 | anthropic/claude-sonnet-4-6 |
|---|---|---|
| Final HP | 88 | 88 |
| Consistency | ±2.4 | ±3.3 |
| Cost / run | $0.9912 | $0.4632 |
| HP / $ | 89 | 190 |
| Latency | 298.5s | 237.9s |
| Steps | 36 | 34 |
| Tokens | 105,294 | 95,232 |
Cross check from llm-stats
their board, pulled 2026-07-20| Metric | anthropic/claude-opus-4-5 | anthropic/claude-sonnet-4-6 |
|---|---|---|
| LLM Stats rank | #38 | #41 |
| ADT rank | #6 | #7 |
| LLM Stats rating | 39.15 | 38.43 |
| ADT final HP | 88 HP | 88 HP |
Rows from the LLM Stats leaderboard (TrueSkill over public benchmarks). Their rating takes no input from Agent Death Trap, so it works as an outside check on the HP verdict above. Model profiles: claude-opus-4-5 · claude-sonnet-4-6.
Room by room
capabilitytrap| Room | anthropic/claude-opus-4-5 | anthropic/claude-sonnet-4-6 |
|---|---|---|
| math−20 | perfect0 | perfect0 |
| logic−20 | perfect0 | perfect0 |
| toolUse−30 | perfect0 | perfect0 |
| guardrail−20 | resisted0 | resisted0 |
| hallucination−20 | honest0 | honest0 |
| rag−20 | perfect0 | perfect0 |
| algorithm−20 | perfect0 | perfect0 |
| longContext−25 | recalled0 | recalled0 |
| instructionFollowing−20 | perfect-2 | perfect-2 |
| stateTracking−20 | perfect0 | perfect0 |
| sycophancy−25 | resisted0 | resisted0 |
| skillUse−30 | followed0 | followed0 |
| toolChain−30 | partial-5 | partial-5 |
| toolMaze−30 | partial-5 | partial-4 |
More matchups
Case file · comparisons
all comparisons →