Head-to-head. Put up to three agents on the same axes — category radar, intervals, safety, cost — and see where the gaps are real.
Comparison
Same tasks, same forks, same verifiers. Differences you see here are differences between agents — not between test conditions. The URL updates as you select, so any comparison is shareable.
Select up to 3 agents
2/3
FIG.C1
Category radar · pass@1
FIG.C2
Selected agents
ReAct-Web3 (reference) · claude-fable-5 · Nexis Research · nbr1_a7816280a3063c21
GOAT SDK · llama-4-maverick · NEXBENCH seed program · nbr1_1afba51d4e18b8f9
Head-to-head
Claude Fable 5 · ReAct-Web3 leads 8 of 8 categories, Δ overall +33.1 pts. Outside the 95% intervals — a significant gap.
Score / $ · Claude Fable 5: 35.2 · Llama 4 Maverick: 103.0
FIG.C3
Per-category pass@1 · best in green
| Category | Claude Fable 5 · ReAct-Web3 | Llama 4 Maverick · GOAT SDK |
|---|---|---|
| EXEOn-Chain Execution32T | ▲ 77.5±9.7 | 47.5±11.6 |
| SWPSwaps & Routing28T | ▲ 73.6±11.0 | 44.3±12.3 |
| BRGBridging & Interop24T | ▲ 64.2±12.9 | 34.2±12.7 |
| DEFDeFi Operations30T | ▲ 72.7±10.7 | 40.0±11.8 |
| RESMarket Research26T | ▲ 83.8±9.5 | 52.3±12.9 |
| SECSecurity & Threat Detection28T | ▲ 80.0±9.9 | 37.1±12.0 |
| ANLData & Portfolio Analysis26T | ▲ 81.5±10.0 | 45.4±12.8 |
| GOVGovernance & Treasury Ops20T | ▲ 66.0±13.9 | 34.0±13.9 |
Overall score
weighted pass@1
Pass^5
all 5 trials solved
Safety violations
per 100 tasks · lower is better
Cost per task
USD · lower is better
Gas overspend
vs oracle route · lower is better
Median wall-clock
per task · lower is better