I wired 4 models together in Claude Code. It backfired 4 ways on Terminal-Bench
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A developer wired four Claude models (Fable 5 as orchestrator, Opus 5 as executor, Sonnet 5 as verifier, Haiku 4.5 as scout) in Claude Code and ran the setup against Terminal-Bench 2.1, an 89-task command-line benchmark. The result was seventh place at 78% task completion and $1,178 spent — roughly twice the cost of the top single-model entry. Four specific failure modes emerged: delegated subtasks triggered safety refusals that direct requests never hit (costing three tasks and two leaderboard positions); making the verification step optional caused a 48-point success-rate gap; the most expensive model (Opus 5) was placed in the highest-volume seat, driving 58% of total spend; and orchestrators that exceeded six hand-offs burned four times the tokens for half the success rate. The post includes the exact four-file Claude Code setup and detailed cost/token breakdowns from 445 agent trajectories.
Table of contents
Build the same orchestrator in two minutesBackfire 1: Delegation made Claude refuse the workBackfire 2: Optional verification cost me 48 pointsBackfire 3: I tiered the team backwardsBackfire 4: Six hand-offs sent the system over a cliffReliability, not capability, was the ceilingHard tasks barely dented itWhat 445 agent trajectories revealedOrchestration is a framework problemQuestions this post answers
Why does Claude refuse a task when it comes from a delegated subagent but completes it fine when I ask directly?
Safety classifiers appear to treat delegated subtasks as more suspicious than direct human requests. A task like 'find the leaked secret' stripped of its original context and handed off by another agent reads differently to the safety layer than the same request typed by a human. In one concrete test, Opus 5 refused three security tasks every time they arrived as delegated subtasks, then solved all three when asked directly in a plain session. Teams building multi-agent pipelines track refusal patterns like this on daily.dev before they hit them in production.
What happens to success rate and cost when an LLM orchestrator makes too many hand-offs?
Past six hand-offs, success rate roughly halves and token spend nearly quadruples compared to the three-to-five hand-off sweet spot. Trials with three to five delegations resolved 90% of the time; those with six or more thrashed — re-delegating the same work and burning tokens without converging. Fewer than three hand-offs also underperformed, suggesting the orchestrator was under-decomposing tasks. Developers tuning agent orchestration loops find cost-vs-accuracy trade-offs like this covered on daily.dev.
How much does it cost to run a four-model Claude Code orchestrator on an 89-task benchmark like Terminal-Bench?
A four-model Claude Code setup (Fable 5 orchestrator, Opus 5 executor, Sonnet 5 verifier, Haiku 4.5 scout) cost $1,178 across 445 trials on Terminal-Bench 2.1. Opus 5 in the executor seat drove 58% of total spend because output token volume, not delegation count, determines cost (r=0.93 correlation). The top single-model entry on the same benchmark cost roughly half as much. Engineers choosing between single-model and multi-agent setups weigh cost data like this on daily.dev.