I interviewed Gemini for a job this week. I told it so, in those words. Then I gave it the one condition: it absolutely could not mess up my existing code. Do no harm. Before it touched anything, I asked it for a plan on exactly that front.
It did that well. Not that I understood most of it. Then I let it examine the code.
In a day working with Gemini, I’ve accomplished more than a week with Codex and Fable working side by side. I’m very impressed. I’m also keeping my head, because the model came out two days ago, and two days is an audition, not a track record.
Some context on what I’m actually running, for anyone doing the same.
The new arrival is Gemini 3.8 Flash, which Google shipped on September 2 and calls “our most intelligent workhorse model”, with gains over 3.7 Flash “across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains.” Google’s coding claim is that it “outperforms most larger frontier models in autonomously solving complex engineering problems end to end, only at a fraction of the cost.” The introductory price is $0.75 per million input tokens and $3.75 per million output, good through December 31. It arrived three weeks after 3.7 Flash. Google is shipping at a pace that makes every verdict provisional, mine included.
Fable is Anthropic’s model. Fable 5 came out in June as what Anthropic called “a Mythos-class model that we’ve made safe for general use”, and the 5.1 update landed on September 1, with Anthropic reporting that testers found it “especially skilled at verifying its own work”. Codex is OpenAI’s coding agent. The open-source CLI came out in April 2025 and runs locally on your computer, in the terminal, against your own files.
So here is the division of labor as of today. I’ll change it the moment it stops earning its keep.
Codex leads. Fable architects. Gemini 3.8 builds. Codex double-checks the work and merges when Claude concurs.
Notice the last clause. Nothing merges on one model’s say-so, and the model that wrote the code doesn’t get the final vote on whether it ships. The new hire builds. Two models I’ve worked with longer, from two different companies, have to agree before anything lands.
The open question is whether the do-no-harm plan holds once the code gets ugly and the tasks get long. Two days in, Gemini has the job. Keeping it is another matter.
