Yet another batch of meaningless autonomous agent benchmarks :)

https://github.com/korchasa/flowai-experiments/tree/main/agents-comparison

Burned 40% of my limit running benchmarks close to my real tasks on opus/fable/gpt-5.5 — fully autonomous agent work: app generation from scratch, a project audit, and three implementation tasks of varying difficulty.

What can be said with at least some confidence:

• fable beats opus-4.8 and gpt-5.5 on result quality. My working hypothesis: fable medium = opus xhigh.
• opus xhigh is unexpectedly expensive due to overly long reasoning. Sometimes more expensive than fable.
• Looks are still a pain. Everything is dark-neon-identical.
• Proper testing will take 1-2 weekly limits on claude x20.

Hypotheses:

• The best model choice will depend on the project’s stage of development.
• In some cases, more expensive but higher-quality models can be justified even on cost over the span of a single task, without counting technical debt.
• Beyond some threshold, longer reasoning no longer improves quality — it only increases cost.