Published: 2026-07-14
Analysis & perspective

Which AI Model Should You Pick? Nate B Jones on GPT-5.6 vs Fable 5

Chapters / key moments (click to jump — plays here on the page)

Nate B Jones argues the right way to choose between GPT-5.6 and Fable 5 isn't benchmark scores — even his own — but your own working style: how you prompt, how you think, and how you actually get to your best work. He explains why he reaches for GPT-5.6 (plus Codex) for his high-intent, verbose prompting, while acknowledging Fable 5 is the stronger generalist for high-level, ambiguous work.

Source video

"Your Next AI Subscription Shouldn't Be ChatGPT 5.6 Or Fable 5. It Should Be Both." by Nate B Jones — Watch on YouTube →

Key Takeaways

Don't lead with benchmarks. Jones benchmarks these models on a private suite, but his advice is to pick by looking at your own best work and the process you use to get there — then ask which model accelerates that loop.
GPT-5.6 "Soul" is a specialist. OpenAI has invested in reinforcement learning on existing model lineages, making it very strong on knowledge work and long-running agentic coding — but, for him, it lacks the "big model smell."
Fable 5 is the generalist. Anthropic's pre-training investment gives it a stronger ability to generalize — better for high-level ambiguity, wrestling with concepts, and front-end instinct.
His workflow fits GPT-5.6. He dictates lengthy, specific prompts via Whisper Flow and steers Codex, which learns from his work to improve its skills — a high-intent loop that suits a persistent, instruction-following model.
Cheaper coding options exist. For fast, efficient coding he points to OpenAI's Luna series, Grok, or GLM 5.2 — and to orchestrating cheap models under a smart architect (his "Ringer" setup, with Fable 5 as the architect breaking down intent).

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

His selection method: start from your work, not the leaderboard
Look at your best work and the process you use to get there, then ask which model speeds up that loop. He says his own benchmark scores should not be your deciding factor.
When GPT-5.6 Sol (transcribed "Soul") fits, per the speaker
You write long, specific, somewhat technical prompts (he dictates them with a voice tool), want explicit instructions followed persistently, and work in Codex, where he says the harness lets the agent learn from what you do and improve skills. He reports Sol scoring 93 on his own knowledge-work test, which he calls Dingo.
When Fable 5 fits
Your prompts are higher-level or ambiguous and you want the model to infer intent; he also credits it with good front-end taste. He still uses Fable as the architect in his Ringer setup, where one central model orchestrates cheaper worker models.
Cheaper alternatives he names for coding
OpenAI's Luna series (released with Sol and Terra), Grok, and GLM 5.2.

Gotchas

Analysis video, not a how-to: the recommendations are one person's experience. The "model families" idea (OpenAI tuned by reinforcement learning on existing lineages and strong at explicit long-run coding; Anthropic investing in pretraining and better at ambiguity) is his interpretation.
He argues benchmarks, including his own, do not capture what these models are like to work with, and says to test them on your own workflows.
He is critical of ChatGPT Work and Claude Cowork as tools designed by engineers for non-engineers, and says knowledge-work harnesses are still weak.
The model-picker tool he links is not described in the transcript.