Descriptions:
David Ondrej provides a day-zero first look at GPT-5.5 within minutes of its public release, framing the model as OpenAI’s direct answer to Anthropic’s as-yet-unreleased Claude Mythos. Unlike Mythos, GPT-5.5 is immediately available to ChatGPT Pro users and inside Codex, and Ondrej dives straight into benchmarks and live testing: the model scores 82.7 on TerminalBench and 56.6 on SWE-bench Pro using real GitHub issues, alongside claimed improvements in understanding user intent and generating higher-quality output with fewer tokens than GPT-5.4.
The hands-on portion centers on Codex, where Ondrej tests GPT-5.5’s ability to recreate a complex UI from a single screenshot—a unicorn-themed graphic—and finds the reproduction nearly identical without manual correction. He then chains image generation, computer use, and Codex’s CLI skill together to attempt building a Doom-style macOS game from scratch, leveraging the newly updated Codex interface that now includes a built-in browser preview window.
Ondrej also presents a competitive analysis arguing that Anthropic is facing a compute shortage: users are reporting instruction-following regressions in Claude Opus 4.7 compared to 4.6, usage limits are tightening, and Mythos remains unshipped. He notes he currently defaults to Claude Opus 4.6 in Claude Code because 4.7 is less reliable for instruction-following, and suggests that if GPT-5.5 delivers on its agentic coding claims, 2026 could mark a meaningful shift in the OpenAI-Anthropic competitive balance.
📺 Source: David Ondrej · Published April 23, 2026
🏷️ Format: Review
