GPT-6.1 Terra Slices API Costs by Half, but Community Benchmarks Show a Quality Gap
The price drop makes agentic automation affordable enough to route more small tasks to AI, but the quality trade-off means complex, high-stakes builds still need costlier models like Sol or Astra.
OpenAI's GPT-6.1 Terra now runs in ChatGPT Work and Codex, with API prices halved: input drops from $2 to $1 per million tokens, output from $12 to $5, and cached input at $0.05. In a side-by-side Belem Tower rebuild, the new model took roughly 4x longer and 48% more tokens than GPT-5.6 Terra, yet the bill came to $0.047 vs $0.077, a 38% saving. Artificial Analysis reports a 6-point score gain in its composite test, with average task cost down from $1.40 to $0.40.
Community users pushed the model on 3D modeling and web pages. A crab build finished in 5:38, a parrot riding a bike in 2 minutes (but static), and a T-Rex motorcycle HTML page in Fast mode. A voxel island via Three.js took 9:13, but the user said overall quality and detail still fall short of Astra, which had taken over 40 minutes. Another user tried an Interstellar black hole and found the result unconvincing, saying GPT-series always struggles with that effect.
OpenAI still recommends Sol and Astra for the most complex programming. On Artificial Analysis's code agent index, Terra scores 48 at max, vs Sol's 54 and Luna's 31. The hard-guess rate in AA-Omniscience fell from 88% to 38%, suggesting better calibration. Terra supports 1.05M token context, 128K max output, and reasoning tiers from none to max. API inputs beyond 272K tokens trigger a 2x multiplier on input/cache and 1.5x on output.
Lower per-token pricing doesn't guarantee lower bills when token usage expands; the Belem Tower test shows a 48% token increase can erode much of the per-token saving.
Terra's price drop is paired with a speed penalty: the new model takes ~4x longer on the same rebuild, so teams paying by wall-clock time may not see the full benefit.
OpenAI's own guidance to use Sol and Astra for complex coding acknowledges Terra's quality ceiling; the model is a cost-efficient workhorse, not a top-tier coder.