Terra returns in the 6.1 lineup after skipping the 6 release. It's roughly the same performance as GPT-5.6 Sol at 80% less, but a Best of 5 run shows how your mileage will vary.
Terra is now the best value for performance out of any model, and about $2 and 30 minutes of time will get an approximation of complete B2B software.
Backend scores spread 1.34 points across the five runs, frontend 0.79. R3 led both tracks at 5.59 and 6.74. R2 posted the weakest frontend, 4.80, yet the second-best backend.
But, not only will your mileage vary, but there's no correlation in more tokens leading to better outcomes. Left alone, the model will make and commit to mistakes, and then spend more time dealing with those issues.
R3 cost $1.99 and averaged 6.17. R4 cost the most at $2.83 and averaged 5.52. Higher spend came with lower scores, a −0.58 correlation.
Despite stronger BE than FE performance, there are wild swings in its ability to deliver, and causes me to doubt the model's ability to carry projects on its own.
Accessibility is the weakest frontend dimension, with a median of 3.8 and a best run of 4.5.
Validation and access is the weakest backend dimension at a median of 4.8, with HTTP and schemas next at 5.0.