Summary
This review looks at OpenAI’s newly released 6.2 Astra and 6.2 Soul models and asks whether they close the gap with Anthropic’s Opus 5.5 and Fable models in AI coding. Theo reports strong results in his own benchmarks, with Astra leading across the board and Soul a close second at far lower cost. Running his benchmark suite on Astra cost roughly half as much as on Fable 5.2, and Soul came in near an eighth.
Despite the numbers, Theo explains why he finds the models frustrating to use day to day. He points to unpolished, generic-looking UIs, odd interface behavior, and rough edges in generated games, while praising the graphics quality of 3D output. He also discusses his experience with safety blocks, tests in real games, and a letter he wrote to OpenAI during early access.
On pricing, 6.2 Soul matches 6.1 Soul, while Astra gets cheaper: cached input drops to $8 per million tokens with a 95% discount and output falls from $50 to $40 per million. The video is useful for developers deciding whether to move workloads off Claude.
📺 Source: Theo – t3․gg · Published November 18, 2026
🏷️ Format: Review







