---
id: "lib-2026-11-04-tested-the-popular-idea-opus-plans-and-reviews-a"
type: "library"
title: "Tested the popular idea: Opus plans and reviews, a cheap model writes the code (Проверил популярную идею: Opus думает и ревьюит, а код пишет дипсик или квен)"
lang: "RU"
org: "Ibragim Badertdinov"
source_type: "Practitioner post"
date: "2026-11-04"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-04-tested-the-popular-idea-opus-plans-and-reviews-a/"
original_url: "https://t.me/dealerAI/1996"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Short practitioner post from the Russian Telegram channel \"commit history\" (forwarded by Dealer.AI) with a small but reproducible test: 12 public benchmark tasks published in Harbor and a linked full results table."
summary_basis: "full-text"
original_minutes: "2"
tags: ["ai technology", "token cost", "coding agents", "model routing"]
---

# Tested the popular idea: Opus plans and reviews, a cheap model writes the code (Проверил популярную идею: Opus думает и ревьюит, а код пишет дипсик или квен)

Tier B (Useful, limited): Short practitioner post from the Russian Telegram channel "commit history" (forwarded by Dealer.AI) with a small but reproducible test: 12 public benchmark tasks published in Harbor and a linked full results table.

Summary based on the full text.

## In short

A Telegram post (4 November 2026, forwarded by the Dealer.AI channel from "commit history") tests the popular split where Opus plans and reviews while a cheap model (DeepSeek or Qwen) writes the code. On 12 hard ML and software tasks the split was both worse and more expensive than Opus alone. The author traces the extra cost to prompt caching: the Opus cache expires while the cheap model works, and rebuilding it costs more than the cheap model's whole contribution.

## Key numbers

- 81%, $1.20: score and cost per task for Opus working alone
- 67%, $1.29: score and cost per task for Opus planning and reviewing with DeepSeek Flash writing the code
- $0.71: cost per task when Opus was told its tokens cost 20x more, at the same score
- 78% / 75%, $0.17: scores of the cheap models Terra and Sol working alone, each at $0.17 per task
- more than half: share of the benchmark tasks that were unfair as shipped and had to be fixed

## What it reports

- 12 hard tasks taken from open benchmarks, mostly hard ML/SWE, chosen as closest to the author's own work; about 7 model setups were run.
- More than half of the benchmark tasks were unfair as shipped: tests either failed a working solution or passed a weak one. The author fixed them before the runs.
- Opus alone: 81% at $1.20 per task. Opus plus DeepSeek Flash as coder: 67% at $1.29.
- Most of the money goes on the prompt cache, which expires while the cheap model is working.
- A variant inspired by DeepSeek's Super-agent article: Opus was allowed to write code but told its tokens cost 20 times more. Same score, cost nearly halved to $0.71.
- Cheap models working alone beat every combination: Terra 78%, Sol 75%, both at $0.17 per task.
- Sol reviewed well; Opus was mediocre as a reviewer. The review step reads diffs and sends work back to the worker if it finds a fault.

## How they got it

The author picked 12 tasks from public benchmarks, did a few trial runs, repaired the task tests, then ran about seven setups. Tasks are published in Harbor for reruns, with a full results table, an X thread and a Reddit post linked. The post itself gives only headline numbers.

## Limits

- 12 tasks and an unstated number of runs per setup; no variance or confidence intervals in the post.
- Results depend on the specific harness and caching behaviour; the post does not describe the orchestration in detail.
- Model names such as Terra and Sol are given without versions or providers.
- Only the summary post was read, not the linked full table.

## Why it matters for companies

The "expensive planner, cheap coder" pattern is often sold as a cost saver. This test suggests the saving can disappear through cache misses and hand-off overhead, and that a single cheap model may beat the combination. The finding that over half the benchmark tasks had broken tests is a warning for anyone choosing models from public leaderboards. Run your own small task set, check the tests first, and price the whole pipeline, cache included.

## Key passages in translation

> Больше половины задач в бенчмарках не честные: тесты либо валили рабочее решение, либо пропускали слабое.

More than half of the benchmark tasks are not fair: the tests either failed a working solution or passed a weak one.

> Деньги в основном уходят на кеш промпта. кеш у Opus протухает, пока дешевая модель работает и пересобирать его дороже, чем вся работа дешёвой модели.

The money goes mostly on the prompt cache. The Opus cache goes stale while the cheap model works, and rebuilding it costs more than all of the cheap model's work.

> Дешёвые модели в одиночку обогнали все связки

The cheap models on their own beat every combination.

Original: https://t.me/dealerAI/1996