I tried to make Opus 5.5 cheaper with a cheap coding subagent, but It got worse and more expensive
retained text
**TL;DR:** On 12 hard coding tasks, Opus alone solved 81% at $1.20/task. Opus + DeepSeek V4.1 Flash solved 67% at $1.29. I tried 7 team designs (41 setups, \~1,200 runs). None beat Opus alone, and most cost more. **Why I did this.** Opus is great but expensive. I was looking for one of two things: a team that beats a single model without a big jump in price, or a setup that's much cheaper than Opus and almost as good. So I had Opus plan and review while cheaper models wrote the code. **The tasks.** I used to build small benchmarks from my own agent sessions, but those tasks turned out too easy. This time I picked 12 hard tasks that look like my daily work, from TB4, DeepSWE, SWE-rebench, SWE-bench Pro V2 and Longtail-SWE. Hidden tests decide pass or fail, and every setup ran each task 3 times. I also went through the runs by hand: fixed broken tests or prompts in 9 tasks, and gave zero to "solutions" that just downloaded the real fix from upstream. **The chart** shows the main setups. The best Opus-led team got 69%. Single cheaper models beat every team: GPT-6.1 Terra got 78% and Sol 75%, both at $0.17/task. **A few things stood out:** * A good plan doesn't save a weak worker. Most of the time the team failed, the worker couldn't solve that task on its own either. * Handoffs can break things. In one task, Opus made up a rule that wasn't in the prompt, DeepSeek followed it, and Opus approved the code. The hidden test failed it, even though each model solves this task alone every time. * Waiting is expensive. While the worker codes, Opus's prompt cache expires, and Opus pays to rebuild it. That alone cost more than the whole worker. * How you ask matters. Telling Opus "your tokens cost 20x more than the worker's" instead of banning it from writing code gave the same score at almost half the price. **Cheap models did get better with a reviewer.** Having Sol review their code added 8–11 points for about $0.07/task. But it only worked one way: Sol made Opus's solutions better, while Opus reviewing Sol made things worse. It's only 12 tasks, so small differences are noise, and shorter tasks may look different. Next, I'm testing a 1-hour cache for Opus. Full table, every setup with scores, confidence intervals, cost and time per task. [https://ibragim.dev/leaderboard-v2/](https://ibragim.dev/leaderboard-v2/) If you know a setup that should beat Opus alone, tell me how you'd build it. I'll run it and add it to the table.
document fetched 2026-11-04 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xn0e12/i_tried_to_make_opus_55_cheaper_with_a_cheap