NewsBite analysis · AI, Programming, Tools
Dev.to
A Spent $5,000 On Tokens; So That You Don't Have To (Part 2)
The author recounts spending $5,000 on tokens for AI coding assistants like Claude Code, OpenAI Codex, and Gemini over the course of a month. He describes burning through a $200 Max subscription in the first week, forcing him to use multiple tools to maintain productivity. He criticizes Claude Code as the least reliable tool, preferring to use it only after exhausting rate limits on other tools. He patches open-source Codex CLI to run Claude and Gemini alongside GPT, allowing him to delegate work between agents. The author concludes that current models are overconfident and unreliable, often requiring heavy supervision. This experience highlights growing pains in the AI coding assistant space: rising costs, tighter rate limits, and inconsistent model performance. The broader implication is that developers and teams must carefully evaluate tooling strategies, possibly adopting multi-model workflows to mitigate these issues.
MonitorConfidence: MediumShort term horizon
Published 12 May 2026, 00:50 UTC4 min read
Visit original article ↗Lead image — Dev.to
Summary
TL;DR
- Spent $5,000 on AI coding tool tokens in one month.
- Found Claude Code the least reliable tool among those used.
- Patched OpenAI Codex CLI to run Claude and Gemini models.
- Uses multiple tools to manage rate limits and maintain throughput.
- Believes mixing models yields better results than relying on a single tool.
- Describes current models as 'dangerous idiots when not closely supervised'.
Interesting facts
- The author burns through a $200 Max sub in the first week, needing two Max subs for a full day.
- He promotes latest models to 'product managers' because they are too erratic to write code.