Detailed Analysis
This article centers on an OpenAI product—referred to variously as "GPT 5.6 Sol," "Soul," or "Saul" throughout the transcript (likely a mis-transcription of a model name, possibly "GPT-5.6" running under a "Codex"-like agentic system)—rather than any Anthropic or Claude product. The creator, Nate, demonstrates the model's Ultra tier orchestrating a four-agent workflow to autonomously research, script, generate voice via ElevenLabs, animate an avatar via HeyGen ("Hen"), and edit via Hyperframes, all from a single prompt. The piece is framed as a real-world stress test of agentic video production rather than a benchmark exercise, and it reports strong results: 91.9% on Terminal-Bench 2.1 and 92.2% on BrowseComp, alongside a small-sample internal test where the model won or tied 12 of 13 tasks against a comparison system.
The one direct connection to Anthropic comes late in the piece, when the creator notes that "GPT 5.6 Soul is similarly priced to Opus 4.8" when comparing API billing costs. This offhand remark is significant because it situates the new OpenAI agentic tier within the same competitive pricing bracket as Anthropic's Opus line, Claude's flagship reasoning model. Cost comparisons like this matter increasingly as agentic workflows—where a single user prompt can spawn nine or more sub-agents consuming hundreds of millions of tokens—turn API pricing into a first-order product consideration rather than an afterthought. The creator's finding that a single video run cost roughly $300 in raw token spend, and that "Ultra" mode may have caused the model to "overthink" and over-delegate relative to a "high" effort setting, echoes a broader industry conversation happening around Claude's own extended-thinking and agentic tool-use modes: more autonomy and more parallel agents can multiply cost without proportionally improving output quality, making effort-tuning a practical lever for developers and creators alike.
More broadly, this article reflects the current moment in AI development where foundation model labs—OpenAI, Anthropic, Google—are converging on similar agentic architectures: orchestrator models that spin up specialized sub-agents, chain together third-party APIs (voice cloning, avatar generation, video rendering), and self-verify their own output through automated review passes before returning a finished artifact rather than a chat response. Anthropic's Claude models, particularly Opus and its computer-use and agentic tool-calling capabilities, are widely used in exactly this kind of multi-tool orchestration, and the pricing comparison here signals that Claude Opus remains a benchmark reference point competitors measure themselves against, even in creator-economy contexts far removed from coding or enterprise use cases. The fact that a YouTube creator casually cites Opus 4.8 pricing as a familiar cost anchor illustrates how deeply Claude's pricing and capability tiers have become embedded in the mental models of technical practitioners evaluating any new agentic AI release, regardless of which lab produced it.
Read original article →