How much did Gemini Code cost in one day?
A solo developer says a single day of work with Google's Gemini Code command-line tool cost him $800 in API charges. The session ran for nearly 12 hours as the tool repeatedly tried, and failed, to fix its own errors.
The developer described the episode in a detailed post. He had been building an automated radio broadcast system that involved refactoring playback handlers and wiring up text-to-speech to reproduce a specific announcer's voice. He had been using Claude Code, with its Opus model as his daily driver, and decided to try Gemini Code because his pipeline code needed upgrades for newer model endpoints anyway.
Does a consumer AI Pro plan cover Gemini Code API usage?
He had expected his consumer AI Pro plan to cover access to newer model tiers, or at least to offer discounts. It did not. API token billing, he found, is handled separately, and he would pay for usage by the token.
At first, he was impressed. He wrote that the speed "felt incredible" out of the gate. He also noted that the command-line interface looks almost identical to Claude Code, with the same clean terminal look, diff layout and interactive prompt. He warned that the resemblance is only on the surface, and that the two tools do not work the same way under the hood.
The trouble started on the second day. He asked Gemini in the browser to plan a redesign of a dashboard for the radio station. The browser tool produced a plan with before-and-after mockups. The actual build was simple: static HTML and JavaScript handling WebSocket data streams that arrived every quarter second. He then asked the browser version to write a prompt for Gemini Code, so that nothing would be lost in translation. According to his account, the command-line tool choked on that prompt.
Over the next session, which he said lasted nearly 12 hours, Gemini Code repeatedly brute-forced test failures and tool runs. By the end of the day, the bill came to $800.
The developer's explanation for the problem is a design flaw he attributes to the tool's architecture. He describes Gemini Code as a stubborn, single-threaded agent that rarely runs tasks in parallel. Claude Code, in his telling, breaks large jobs into smaller ones and uses sub-agents to verify results or work through roadblocks. When Gemini Code got stuck, he said, it did not stop to reconsider the design. It rewrote tests, kept hammering at failing fixes and kept making API calls until it hit his spending limit.
He said he then spent the next eight hours using Claude Code to clean up what Gemini Code had left behind.
How can developers avoid runaway Gemini Code costs?
His takeaways were practical rather than sweeping. He found the tool fast for small, isolated scripts and single-file edits. Stateful, multi-file refactors were another matter. He recommends setting strict spending caps before running the tool, and he cautions against assuming that browser prompts will translate cleanly to command-line execution.
The figures and account come from the developer alone and have not been independently verified.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.