Summary
Theo of t3.gg shares his early-access impressions of GPT 6.1 Soul, OpenAI’s newest model, arriving only about a week after GPT 6.0 Soul and just after Anthropic’s Opus 5.5 and Sonnet 5.5 launches.
Rather than leaning on published charts, Theo runs Terminal Bench himself across low, high, extra-high, and max reasoning settings, and explains why he has become more skeptical of benchmarks. He then tests the model on real codebases. In deep code-review audits of his orchestrator project, GPT 6.1 Soul scored nearly the same as Astra at roughly half the cost ($297 versus $584). In a task to find improvements for T3 Code, it scored 87.4, ahead of Astra at 83.8 and Grok 4.7 at 80.7, at about $2.15 compared with roughly $5 for Opus 5.5 and nearly $9 for Sonnet 5.5.
Theo describes the model as reliable and pleasantly boring, with weaker peaks than Opus on edge-of-capability tasks. He also discusses its strong computer-use ability and why the upcoming Ultrafast mode could make it even more compelling. The video includes a sponsor segment for Depot CI.
📺 Source: Theo – t3․gg · Published September 29, 2026
🏷️ Format: Review
