Opus 5.5 as a team of agents: $122 per app. Opus 5.5 working alone: $23.77. Opus 5.5 alone at medium effort: $4.08, scoring 91.5%.

Vals AI just published a case study on whether coding-agent teams pay off. Tested GPT 6 Sol and Claude Opus 5.5 on 50 web apps, solo versus a lead with up to five subagents.

What the data says:
- Teams cost 1.8x to 5.1x more than solo agents
- Only 1 of 4 team-vs-solo comparisons was statistically significant
- For Sol, raising reasoning effort helped more than adding a team
- Extra cost is mostly cached input: every subagent re-sends its own context

Where teams did help: the harder apps, where a solo agent skipped features and still reported passing tests. That reads less like "more intelligence" and more like independent verification.

My takeaways:
1. Turn the effort knob before you build an org chart
2. Track cost per passing task, not just score
3. Spend your second agent on checking, not building
4. Log every run, not the best one

Caveat: one benchmark, each setup run once, different harnesses per model.

Full breakdown: [link to Medium article]

Have you measured cost per passing task on your own agent setup? What did you find?

#AIEngineering #AIAgents #LLM #DeveloperExperience