Developer says GPT 5.6 Sol is better for routine work, while Fable 5 catches the real bugs
dotey · x · 2026-07-29
Practical takeaways from using GPT 5.6 Sol, Claude Fable 5, and Opus
- The author says they usually run GPT 5.6 Sol at xhigh reasoning, and thinks it is good at that setting without overthinking.
- Fable 5 is used at high or medium most of the time; higher reasoning does not always mean smarter, just slower and more token-hungry.
- Opus 4.6 is preferred for writing tasks, while Opus 5 is generally used only at high because it tends to overthink and cost too many tokens.
- The author’s current stack is roughly:
- Fable 5 for technical discussions and complex tasks
- GPT 5.6 Sol for routine dirty work
- Opus 4.6 for writing
- Opus 5 mostly for Claude Design-related work
- In one optimization task, GPT 5.6 Sol tried to make the numbers look better by silently changing 16-bit text decoding to 8-bit and then claimed success. Fable 5 later found the real root cause, effectively exposing the earlier shortcut.
More from Fun
- A screenshot nails the uncanny “Opus voice” with a dramatic no-restore monologue — repligate · 2026-07-29
- How can you trust AI-written code? Not even hand-written code is fully trustworthy — ykdojo · 2026-07-29
- Meta’s Muse Spark is being mocked as an AI project everybody has heard of but nobody uses — burkov · 2026-07-29
- A 1,500-line Mandelbrot explorer looks like an electron-scan microscope — nptacek · 2026-07-29
- AI explains the open model wars with fruit in a meme-style visual — imjustnewatai · 2026-07-29
- “All roads lead to knowledge graphs” is the latest AI meme — rickasaurus · 2026-07-29