One Prompt, Two Agents: Codex's Sub-Agents Out-Build Claude at Vibecoding

Four years ago, Simon Willison generated a silly game premise with GPT-3 and DALL-E: you and a crew of thieving raccoons pull off heists, from bank jobs to art theft. It sat unused until last week, when he decided to actually build it — twice — as a benchmark for where agentic coding tools stand in mid-2026.

Two agents, one prompt

On August 5, Willison fed the premise to Claude Fable 5 running in Claude Code for web and got a working game in one shot: a single raccoon collecting coins and fish in a backyard. Two days later, he ran the identical prompt through Codex Desktop on GPT-5.6 Sol Ultra — a mode, as Willison describes it, where the model "makes aggressive use of sub-agents." The result, Moonlight & Mayhem, is a noticeably more complete game: the player stacks up with two rescued raccoon crewmates inside a museum to extract a golden sardine, closer in scope to the original "series of daring heists" premise than Fable 5's backyard scene.

This is exactly the kind of apples-to-apples comparison that's hard to find in AI coding coverage, which is usually either vendor benchmark theater or vague vibes. Same prompt, same class of task, two different agent architectures — and Willison logged the results in public, including a bug along the way (an early build gave the raccoons oversized, floating eyeball spheres, fixed through a couple more rounds of prompting).

The number that matters: sub-agent orchestration, not raw model IQ

The headline isn't "GPT-5.6 beats Claude." It's that the *harness* — how a coding agent decomposes a build into parallel sub-agent work — visibly changed the output for the same underlying task and premise. Sol Ultra's aggressive sub-agent delegation let it cover more scope (asset generation, multiple game mechanics, crewmate AI) in roughly the same wall-clock window a single-threaded agent used to build something smaller. That's the pattern we watch for in vibecoding: not which foundation model wins a leaderboard, but which agent runtime decomposes a spec into independently-workable pieces without losing coherence.

Codex Desktop's build also used OpenAI's gpt-image-2 to generate its own textures and sprite prompts, folding asset creation into the same agentic loop as the code — one less handoff to a human designer, and one more place where a sub-agent's output has to be checked, not assumed correct.

Willison published the receipt

What sets this write-up apart from typical "AI built my app" posts is the accounting: 52 minutes, 700.7K input tokens, 32.5M cached tokens, 148K output tokens, roughly $23.28 in API cost. For anyone evaluating agentic coding tools for real work, that kind of line-item transparency is more useful than a demo video — it lets you actually budget a sub-agent-heavy workflow instead of guessing.

Why this matters beyond a raccoon game

A throwaway game prompt is a low-stakes way to stress-test something that matters a lot in production: whether an agent's sub-agent fan-out produces genuinely more capable output, or just more output. The eyeball bug is the tell — sub-agents that work in parallel can each ship something individually plausible that doesn't compose cleanly until a human (or another review pass) catches it. That's the same failure mode that shows up in agentic PRs on real codebases, just lower stakes here.

• Same prompt, two runtimes is a more honest comparison than vendor-published benchmarks.
• Sub-agent aggressiveness is now a meaningful, testable variable — not just model choice.
• Published token/cost breakdowns should become standard practice for anyone writing up agentic coding results.

Codex Desktop running GPT-5.6 Sol Ultra — the mode where Sol makes aggressive use of sub-agents — produced a much better game.

— Simon Willison