Descriptions:
Matthew Berman delivers a hands-on review of GPT-5.6, arguing it represents a more substantial upgrade than its version number suggests — describing it as OpenAI squeezing maximum capability out of the GPT-5 pre-training run. The review centers on two extended agentic builds run inside Codex: a full-featured Excel clone that ran autonomously for five days (with the model using computer use to reference the actual Excel application as a guide), and a Minecraft clone that ran for seven days, producing detailed biomes, working inventory systems, and mob behaviors.
Berman includes benchmark data from Box AI’s enterprise-grade evaluation suite, which tests real knowledge work including document analysis, number reconciliation, and expert output review. GPT-5.6 Soul outperformed GPT-5.5 (63.3% baseline) across public sector, life sciences, and healthcare subsets. On pricing, Soul is $5 per million input tokens and $30 per million output — roughly half of Fable 5’s $10/$50 — and Berman emphasizes that Soul also tends to consume fewer tokens to reach the same result, making the practical cost gap even wider.
The review closes with a memorable framing: GPT-5.6 feels like a veteran athlete — high game IQ, efficient, low error rate — while Fable 5 feels like a high-ceiling rookie with more raw potential but less refined execution. For teams running production agentic workflows today, Berman’s conclusion is that GPT-5.6 Soul delivers strong value at a meaningfully lower cost than the current Claude alternative.
📺 Source: Matthew Berman · Published July 09, 2026
🏷️ Format: Review
