Author(s): Lakshman Sai

Originally published on Towards AI.

How to Make Claude Code as Cheap Per Task as GPT-6 Astra, Without Talking Like a Caveman

Astra wins on cost because it writes less. So thousands of developers taught Claude to grunt. The caveman skill’s own author then measured it on real coding work: 8.5%. Here’s where Claude Code’s tokens actually go and the skills that move the number.

After comparing Claude Code and GPT-6 Astra, the article argues that “write less” is only part of why Astra is cheap. In agentic coding, most cost comes from what the model must repeatedly re-read each turn—system prompts, CLAUDE.md, tool catalogs, skill content, MCP definitions, and tool results—so compressing chat-style output won’t fully address costs unless the “reading” overhead is reduced too. The author highlights that Caveman’s demo headline (token reduction) does not translate directly to real coding agents (paired tests show only ~8.5% fewer output tokens, and the skill can add input overhead), because much of the expensive context is outside Caveman’s scope. The rest of the piece lays out how to find real token hotspots (using /usage and /context, plus baseline measurements), and then how to cut context-heavy costs with targeted interventions: stopping log floods via Context Mode, using caveman-compress and on-demand skills to shrink always-loaded files, delegating with cheaper subagents, trimming MCP tool catalogs, and even considering routing/tiering strategies—while warning to test each change incrementally and uninstall skills that don’t actually move your measured numbers.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI