Descriptions:
Matthew Berman hosts a live stream on the day of GPT-5.6’s public release, walking viewers through the model family — Soul, Terra, and Luna — while fielding live questions and sharing analysis from his published review. Berman frames GPT-5.6 as the culmination of the GPT-5 pre-training run: a highly polished, efficient model that excels at direct task execution, contrasted with Fable 5 (Claude), which he describes as a generationally new training run with higher long-term ceiling but more exploratory (and expensive) behavior today.
Pricing is covered in detail: Soul is $5 per million input tokens and $30 per million output, versus Fable’s $10/$50 — and Berman argues the effective cost gap is even wider because Soul typically uses fewer tokens to complete the same task. He walks through GDPVal (an OpenAI-created practical knowledge-work benchmark) where Soul scores just below Fable 5, and notes the three-tier model structure gives teams flexibility that single-model providers like Anthropic currently lack.
Berman also introduces an open-source Codex skill he built called ‘GPT-5.6 Relay,’ available on GitHub, which routes planning and orchestration tasks to Soul while offloading implementation to Terra and deployment to Luna — a quota-optimization strategy he’s been testing in production. The broader competitive context is notable: on the same day, xAI released Grok 4.5 and Meta released its Spark model, while Anthropic reset Claude’s rate limits — a multi-front competitive dynamic Berman unpacks with viewer commentary in real time.
📺 Source: Matthew Berman · Published July 09, 2026
🏷️ Format: Livestream
