---
id: "lib-2026-11-02-how-far-can-cheap-models-go-gpt-gemini-and-claude"
type: "library"
title: "How far can cheap models go? GPT, Gemini and Claude on four HR and labour tasks (安いAIモデルにどこまで任せられる？ GPT・Gemini・Claudeを人事・労務の4業務で比較)"
lang: "JA"
org: "Rounda"
source_type: "Practitioner post"
date: "2026-11-02"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-02-how-far-can-cheap-models-go-gpt-gemini-and-claude/"
original_url: "https://zenn.dev/rounda_blog/articles/0e164342608b66"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Soichiro Tokunaga of Rounda publishes a task-level model comparison with per-task scores, costs and test conditions, but it is a small single-run test on the vendor's own platform, written partly to sell it."
summary_basis: "full-text"
original_minutes: "12"
tags: ["ai technology", "enablement", "hr operations", "model selection", "cost"]
---

# How far can cheap models go? GPT, Gemini and Claude on four HR and labour tasks (安いAIモデルにどこまで任せられる？ GPT・Gemini・Claudeを人事・労務の4業務で比較)

Tier B (Useful, limited): Soichiro Tokunaga of Rounda publishes a task-level model comparison with per-task scores, costs and test conditions, but it is a small single-run test on the vendor's own platform, written partly to sell it.

Summary based on the full text.

## In short

A Zenn post (2 November 2026) by Soichiro Tokunaga on the Rounda Tech Blog. It compares GPT-6 Luna and Gemini Flash-Lite (light models), Claude Sonnet 5.5 (mid-tier baseline) and GPT-6 Astra (top-tier reference) on four Japanese HR and payroll tasks. Light models matched the top on reading figures from supplied charts but fell behind on knowledge of current rules. No model knew the October 2026 minimum-wage revision. Sonnet gave the best balance overall; the author argues for routing by task, not one model for everything.

## Key numbers

- 8/8 at about 1/20 the cost: GPT-6 Luna on chart reading, matching Claude Sonnet 5.5 ($0.0026 vs $0.053 for 8 questions)
- 5/10: GPT-6 Luna on statutory-figure questions answered from model knowledge (Sonnet and Astra 8/10)
- about 83% vs 92%: average accuracy of Sonnet vs GPT-6 Astra across four tasks, at about a quarter of the cost ($0.37 vs $1.6)
- 0 models: that knew Tokyo's October 2026 minimum wage (1,280 yen) without web search

## What it reports

- Task 1, reading values from charts and tables (8 questions): Luna 8, Flash-Lite 6, Sonnet 8, Astra 8. Flash-Lite calculated correctly but then rounded both answers to 63 yen and picked the wrong prefecture.
- Task 2, current statutory figures without web search (10 questions): Luna 5, Flash-Lite 1, Sonnet 8, Astra 8. Luna gave last year's employment-insurance rate (5.5/1,000 instead of 5/1,000).
- Task 3, finding 8 planted errors in a part-timer work rule: Luna 5, Flash-Lite 2, Sonnet 6, Astra 7. Only Sonnet and Astra caught last year's health-insurance rate (9.91% vs 9.85%); all missed the outdated minimum wage (1,226 yen). No model flagged a correct passage as wrong.
- Task 4, building a payroll correction sheet from contracts, pay records and an email (4 required items): Luna 2, Flash-Lite 0 (despite 37 fields), Sonnet 3, Astra 4 (about 70 fields). Sonnet missed the minimum-wage shortfall, so its final amount was wrong.
- Average accuracy: Luna 66%, Flash-Lite 28%, Sonnet 83%, Astra 92%. Total test cost: $0.017, $0.036, $0.37, $1.6.
- Flash-Lite sometimes invented figures even when told it could answer "unknown". Astra left uncertain amounts blank with a note to confirm against the latest published figures.
- At 10,000 document checks a month: about $35 (Luna), $30 (Flash-Lite), $480 (Sonnet), $2,400 (Astra).

## How they got it

Run 15-21 October 2026 on Rounda's RoundaChat platform, one run per question, default reasoning effort, no caching, web search off. Questions designed with Gemini, answers and chart images produced in Python and checked by eye. Task 2 was graded by an AI scorer (1-5) as well as by correct count. Costs are API costs shown by the platform at 150 yen to the dollar.

## Limits

- Single runs, so no measure of variance; default effort settings only.
- Small question sets (4 to 10 items per task) built by the vendor; the article promotes RoundaChat.
- Japanese HR rules only; the author says gaps between models will change with task and prompt.

## Why it matters for companies

The result supports routing by task type: when the answer is in the supplied document, the cheapest model may be enough, and paying for a mid-tier model everywhere costs about 20 times more on those tasks. When the task depends on current rules, no model can be trusted without retrieval or a human check, because rule changes after the training cutoff are invisible to it. The HR and payroll errors here are the kind that turn into underpayment claims, so a check against official sources belongs in the process design.

## Key passages in translation

> 改定直後の制度や数値は、AIが知っている前提にせず、人が確認する運用が必要です。

For rules and figures just revised, do not assume the AI knows them; the process needs a person to check.

> 推測で値を埋めてしまう軽量モデルとは対照的に、上位モデルほど不確かな値は人に委ねる傾向がありました。

In contrast to light models that fill values with guesses, the higher-tier models tended to leave uncertain values to a person.

> すべての業務をSonnetに統一してしまうと、Lunaで足りる定型業務にまで約20倍のコストを払い続けることになります。

If you standardise every task on Sonnet, you keep paying about 20 times more even for routine tasks Luna can handle.

Original: https://zenn.dev/rounda_blog/articles/0e164342608b66