Descriptions:
Matthew Berman hosts a live stream covering the GPT-5.5 launch in real time, joined by guest Ray Fernando. The opening segment tracks OpenAI’s pre-release “vague posting” on social media, including a base64-encoded tweet that decodes to “5.5” — a small piece of launch-day lore that captures how the community monitors OpenAI signals before official announcements.
The substantive portion focuses heavily on benchmarks. Berman walks through OpenAI’s published results on Frontier Math Tier 1 and Tier 2, Terminal Bench, and the Artificial Analysis intelligence index — a composite of multiple benchmarks plotted against output token count. The key finding: GPT-5.5 scores higher than GPT-5.4 across most tasks while generating roughly half the output tokens on some benchmarks, making it significantly more token-efficient despite a higher nominal price per token. Claude Opus 4.7 is placed at approximately 79–80% on the Artificial Analysis index, below GPT-5.5’s scores. The concept of “intelligence per token” is used as a framing for evaluating cost-effectiveness.
Berman also walks through official OpenAI demo videos featuring WebGL trajectory rendering from a NASA JPL Horizons prompt. The stream closes with discussion of Codex as the preferred environment for GPT-5.5 coding tasks. Useful for practitioners tracking comparative model performance and API cost dynamics when evaluating GPT-5.5 against Claude Opus 4.7 for production deployment.
📺 Source: Matthew Berman · Published April 23, 2026
🏷️ Format: Livestream
