We Tested GPT 5.6 Sol Early
Summary
Theo and Ben discuss their extensive early access experience with OpenAI’s GPT-5.6 Sol, praising its superior long-running capabilities and spatial reasoning while noting it remains inferior to Anthropic’s Fable. They highlight improvements in complex coding tasks and mobile development, though frontend design remains weak. The hosts compare the model’s literal execution style to Fable’s intuitive orchestration, describing 5.6 as the peak of the current generation. They also share anecdotes of burning over $130,000 on tokens through experimental loops and sub-agents to test the model’s limits.
Topics discussed
Intro: Delayed release and early access context
Ben's introduction and massive API spending
The 'Fable' model vs. GPT-5.6 comparison
Behavioral shifts: Expectations and UI generation issues
3D reasoning and Blender asset generation
iOS development struggles with SwiftUI
Advanced workflows: PRs, Lakebed, and Hermes
Model personality: Overcomplication and testing
Tiering analogy: PS3 vs PS4 generation leap
Naming conventions and Anthropic vs OpenAI strategy
Sponsor segment: Clerk authentication components
Cost analysis: Sub-agents and usage spikes
Computer use: Browser automation and daily tasks
T3 Code integration and sub-agent orchestration
Codex vs Claude Code: Workflow primitives
Deep dive: GPT-5.6 vs Fable 5 capabilities
Conclusion: Long-run reliability and final thoughts
Listen ad-free on Castria