<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr" -->

---
title: Aider, Claude Code, and OpenClaw ran an identical model....
description: Three independent benchmarks — an indie June study, Composio's August comparison, and Artificial Analysis's coding-agent index — found that swapping the...
canonical: https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold. | daily.dev
og:description: Three independent benchmarks — an indie June study, Composio's August comparison, and Artificial Analysis's coding-agent index — found that swapping the...
og:url: https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr
og:image: https://api.daily.dev/og/posts/AizeJzWqR.png
og:image:alt: Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold.
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold.

**[The New Stack](https://daily.dev/sources/newstack)** · 9 min read · 0 upvotes · 0 comments

## Summary

Three independent benchmarks — an indie June study, Composio's August comparison, and Artificial Analysis's coding-agent index — found that swapping the coding-agent harness while holding the model fixed changes token usage and cost by multiples, even up to 70-fold. The main driver is a 'startup tax': the system prompt and tool descriptions a harness resends every turn, ranging from ~700 tokens (Aider architect mode) to ~26,000 (OpenClaw). A second factor is prompt caching: Claude Code drew only 1.5% of input tokens from cache versus ~70% for Codex, largely because Claude Code used the Anthropic-style messages endpoint on OpenRouter, which appeared to reduce cache hits. Pass rates also varied 20 points across harnesses on identical models, and stress tests found some harnesses (Kilo, Opencode) silently truncated 83-89% of injected context while still reporting success. The piece recommends platform teams measure cost per verified outcome, cached share on live traffic, and prompt fidelity under load rather than raw token counts.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://thenewstack.io/agent-harness-token-costs>

## Questions this post answers

### Why does Claude Code cost more per task than other coding agent harnesses using the same model?

Claude Code draws a much smaller share of its input tokens from cache than rivals, around 1.5% versus roughly 70% for Codex and 57% for OMP. Fresh input costs about five times as much as cached input, so even when Claude Code uses fewer raw tokens than a competitor, its bill can be higher because so little of its traffic hits the cache.

_Anyone comparing agent harness costs before standardizing a platform can track findings like this on daily.dev._

### What causes the huge token overhead differences between AI coding agent harnesses like Aider and OpenClaw?

Each harness resends a 'startup tax' of system prompt, tool descriptions, and environment setup on every turn before doing any real work. That floor measured around 700 tokens for Aider in architect mode versus roughly 26,000 tokens for OpenClaw, and multiplying that floor by turn count predicted total tokens per solved task with an R-squared of 0.99 across two different models.

_Teams tuning agent spend can follow harness cost breakdowns like this on daily.dev before picking a tool._

### Does a cheaper AI coding agent harness mean lower task success rates?

Not necessarily, but there is a real tradeoff. Across eight harnesses tested on identical enterprise workflows, pass rates ranged from 46.7% for OpenCode to 66.7% for Pi Agent, a 20-point spread on the same model, showing that harness choice moves cost by multiples and task success by percentage points as separate, unrelated effects.

_Developers weighing cost against reliability across agent harnesses can track these comparisons on daily.dev._

## Similar posts on daily.dev

- [Claude Code Sends 4.7x More Tokens Than OpenCode Before Reading Your Prompt](https://daily.dev/posts/claude-code-sends-4-7x-more-tokens-than-opencode-before-reading-your-prompt-9m02iom1z) · Hacker News · 6 upvotes · 0 comments
- [The Harness Tax: The Dead Weight Inside Your Coding Agent](https://daily.dev/posts/the-harness-tax-the-dead-weight-inside-your-coding-agent-iiu0gkchd) · portkey · 0 upvotes · 0 comments
- [How We Cut Our Claude Code Token Usage 2.8x\!](https://daily.dev/posts/how-we-cut-our-claude-code-token-usage-2-8x--p2fzdwf2l) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments

---

Tags: [#claude-code](https://daily.dev/tags/claude-code), [#ai-inference](https://daily.dev/tags/ai-inference)

[View this post on daily.dev](https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold.","url":"https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr"},"datePublished":"2026-08-27T17:55:26.749Z","dateModified":"2026-08-28T12:17:16.777Z","description":"Three independent benchmarks — an indie June study, Composio's August comparison, and Artificial Analysis's coding-agent index — found that swapping the...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f01245b33c0e0b0e0cf2ca0ca23d6007?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f01245b33c0e0b0e0cf2ca0ca23d6007?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The New Stack","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"The New Stack","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/newstack","url":"https://daily.dev/sources/newstack"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"claude-code,ai-inference","timeRequired":"PT9M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The New Stack","item":"https://daily.dev/sources/newstack"},{"@type":"ListItem","position":3,"name":"Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold."}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/aider-claude-code-and-openclaw-ran-an-identical-model-token-use-varied-70-fold--aizejzwqr#faq","mainEntity":[{"@type":"Question","name":"Why does Claude Code cost more per task than other coding agent harnesses using the same model?","acceptedAnswer":{"@type":"Answer","text":"Claude Code draws a much smaller share of its input tokens from cache than rivals, around 1.5% versus roughly 70% for Codex and 57% for OMP. Fresh input costs about five times as much as cached input, so even when Claude Code uses fewer raw tokens than a competitor, its bill can be higher because so little of its traffic hits the cache. Anyone comparing agent harness costs before standardizing a platform can track findings like this on daily.dev."}},{"@type":"Question","name":"What causes the huge token overhead differences between AI coding agent harnesses like Aider and OpenClaw?","acceptedAnswer":{"@type":"Answer","text":"Each harness resends a 'startup tax' of system prompt, tool descriptions, and environment setup on every turn before doing any real work. That floor measured around 700 tokens for Aider in architect mode versus roughly 26,000 tokens for OpenClaw, and multiplying that floor by turn count predicted total tokens per solved task with an R-squared of 0.99 across two different models. Teams tuning agent spend can follow harness cost breakdowns like this on daily.dev before picking a tool."}},{"@type":"Question","name":"Does a cheaper AI coding agent harness mean lower task success rates?","acceptedAnswer":{"@type":"Answer","text":"Not necessarily, but there is a real tradeoff. Across eight harnesses tested on identical enterprise workflows, pass rates ranged from 46.7% for OpenCode to 66.7% for Pi Agent, a 20-point spread on the same model, showing that harness choice moves cost by multiples and task success by percentage points as separate, unrelated effects. Developers weighing cost against reliability across agent harnesses can track these comparisons on daily.dev."}}]}
```

