Descriptions:
Nate Herk runs structured cost-and-performance experiments comparing GPT 5.5 against Claude Opus 4.7, focusing on the metrics that matter most for production use: token efficiency, task completion speed, autonomous decomposition, and actual dollar cost per job. On TerminalBench 2.0, GPT 5.5 scores 82.7 versus Opus 4.7’s 69.4, and also beats GPT 5.4’s 75.1. OpenAI’s internal Expert Suite and GDPVal benchmarks show GPT 5.5 ahead on knowledge work and frontier math, though SWE-bench Pro — resolving real GitHub issues — still belongs to Opus 4.7.
The pricing picture is nuanced. GPT 5.5 costs $5 per million input tokens and $30 per million output tokens, slightly more expensive per token than Opus 4.7 ($5 input / $25 output). But Herk’s experiments show GPT 5.5 completes equivalent tasks with dramatically fewer tokens — cutting effective cost from roughly $5 to roughly $1 per complex Codeex task — while also running about 3.5x faster (4 minutes vs 14 minutes in one head-to-head). The video also covers GPT 5.5’s improved autonomous decomposition: given a vague prompt with no follow-up questions allowed, GPT 5.5’s first-pass output required significantly less iteration.
Herk also walks through Codeex upgrades bundled with GPT 5.5, including multi-agent parallel execution and reusable workflows. For engineers and AI practitioners deciding whether to switch default models, this video provides the clearest cost-per-task comparison currently available between OpenAI’s and Anthropic’s latest flagships.
📺 Source: Nate Herk | AI Automation · Published April 23, 2026
🏷️ Format: Comparison
