# Claude and Codex Model Atlas

A practical breakdown of how to split tasks across effort levels and models in Claude and Codex. The goal is to preserve strong results without burning through a $200 subscription.

Author: Maxim Tolmachev
Published: 2026-10-08
Language: en
Canonical: https://maximtolmachev.com/blog/claude-and-codex-model-atlas/
Original source: https://t.me/mtolm/106

![Claude and Codex Model Atlas](/blog-images/tg-106.jpg)

Last time, I explained why the drop in the dollar equivalent in Codex isn’t a disaster if you choose models wisely. Today, let’s look at a practical atlas: what to use in Claude and Codex for different types of tasks.

The main lever for managing your quota right now is the effort parameter. It’s not a magic “make it better” switch, but a budget of time and tokens for the model’s internal reasoning. If you crank it to the maximum for a simple task, the model won’t become smarter—it’ll just burn through your limits and make you wait three times longer.

Task allocation atlas: 3 effort levels in Codex and Claude for $200 subscriptions

Based on the latest Artificial Analysis benchmarks in my interactive model economics dashboard, the optimal working setup looks like this:

1️⃣ OpenAI Codex stack

• 50% of tasks — Luna 6 xhigh / max (Score 35–38, $0.04–0.07 per task).
Parsing, tests, small scripts, and cheap sub-agents that do the dirty work for the main model.

• 45% of tasks — GPT-6.1 Sol medium (Score 48, $0.21 per task).
The main workhorse. In terms of intelligence, it has caught up with yesterday’s top models, while completing a task in 2m 38s versus 6m 29s for Luna max—2.5 times faster at the ridiculously low price of 21 cents.

• 5% of tasks — GPT-6.1 Sol max (Score 52, $0.72 per task).
Critical core logic, security, complex math, and deep audits.

👉 Codex workflow result: the average task costs only about $0.16. A $200 subscription comfortably covers more than 1,200 real tasks per month.

2️⃣ Anthropic Claude stack

• 70% of tasks — Claude Haiku 5.5 xhigh (Score 41, $0.12 per task).
According to AA benchmarks, this little monster beat Luna max (41 versus 38) and costs only 12 cents. It’s where you should offload all the routine work, document drafts, and targeted codebase edits.

• 25% of tasks — Claude Opus 5.5 medium (Score 51, $1.34 per task).
Deep product code, complex multi-file refactoring, and architectural decisions.

• 5% of tasks — Claude Opus 5.5 high (Score 54, $1.82 per task).
Maximum reasoning for the trickiest bugs, where every nuance of the logic matters.

👉 Claude workflow result: the average task costs about $0.51, with a high average intelligence score of 44.2.

Why are the ratios different? In Codex, Sol 6.1 is so cheap ($0.21) that it makes sense to keep it enabled for almost half of all tasks. In Claude, Opus costs $1.34–1.82 (6–8 times more than Sol), so Haiku is the workhorse at 70%, while Opus is brought in selectively.

🚫 What I excluded from the pipeline

• Claude Fable 5.1: with comparable intelligence (Score 47–53), a task can cost up to $7.63, plus there’s a cap of up to 50% of the Max plan’s weekly limit.

• GPT-6 Astra: burns through the quota like a sports car ($0.82–3.26 per task) without delivering a quality advantage over Sol 6.1.

You can see the detailed trends and a comparison with the September snapshot in the previous post, or explore all the model data points at maximtolmachev.com/model-economics/.

What effort level do you usually use by default?
