Best value is DeepSeek V4.1 Flash, Best under $1 is Amazon Nova Micro, Best quality stays Claude Opus 5.5. The table
adds Amazon Nova Lite, Amazon Nova Micro, Gemma 3 4B, GPT-4.1 nano and Mistral Nemo. The chart is the redesigned one
from the blog, and the cost line no longer says list prices.
Cost is per 1,000 add requests (the unit on the pricing page) and the model is
described as the one that extracts memories. The rate-limit note is gone and the
images carry the same unit.
Each model's accuracy is now the average of its first run and a second that gives every missed question one more try. The free tier shows its published 92.5%. Seven models are tied at the top, between 92.6% and 93.5%.
Accuracy, cost per 1,000 adds and p95 latency for the 13 models we evaluated on LoCoMo, the three picks, the chart, and the steps to add a key under Settings -> BYOK. Links to the full evaluation on the blog.