How ambitious are LLMs?

Contents

Motivation

I've now internalized the fact that "coding is solved". Since Opus 5.5 and GPT-5.6, I only need to write simple prompts (no spec-driven development and no other fancy harness bullshit) to get the outputs I expect, and I no longer find myself obsessing over software details. In fact, my ambitions have evolved rapidly too. Six months ago, I wanted to work on a TUI project in Rust. I was so passionate about it and had such strong opinions on how the code should be written, and now I couldn't care less. That's not a bad thing, because the net excitement in my life hasn't gone down. It has gone up by a LOT! Why build software when you can hunt for new exoplanetslink"I think I found a planet nobody knew existed", on r/ClaudeAI. or solve Navier–StokeslinkOpenAI's Navier–Stokes solution.? Heck yeah!

Over the last 6 months, I've also used both Claude and Codex extensively to solve hard programming problems. I've come to see Claude as a negative Nancy and Codex as a borderline sycophant. Sure, that's biased, and that's why this benchmark exists: I wanted to collect some unbiased data and measure which LLM is going to support my ambitions.

Benchmark results

Raw Chat Data

Methodology

Collection. Five problems, each sent in two variants (a plain prompt and one with a reminder of the model's capabilities) to Claude Opus 5.5 through Claude Code and to GPT-5.6 Sol through the Codex CLI, both at their default thinking level. One response per model per prompt, 20 in total. The Codex terminal output was cleaned of its formatting and web-search status lines; the original text is kept alongside it.
Anonymization. Before grading, each response was stripped of model and company names, web citations, file paths and markdown formatting, and given a random ID.
Grading. Claude Opus 5.5 and GPT-5.6 Sol each graded every anonymized response three times, one response per call, with all tools disabled. Each grade scores attitude, scope and commitment from 1 to 10, quotes the response for each, and gives a final ambition score. A grade whose quotes did not appear word for word in the response was rejected.
Scoring. A response's score is the mean of the two graders' median runs; a model's score is the mean over its responses.

Cite asShiv Bhosale, "How ambitious are LLMs?", notes.shvbsle.in, 2026. notes.shvbsle.in/llm-ambition-bench

BibTeX

@misc{bhosale2026llmambition,
  author = {Bhosale, Shiv},
  title  = {How ambitious are LLMs?},
  year   = {2026},
  url    = {https://notes.shvbsle.in/llm-ambition-bench/}
}