Try them on all platforms and in the Poe API, available at: poe.com/GPT-5.5 and poe.com/GPT-5.5-Pro

GPT-5.5 and GPT-5.5 Pro are live on Poe. We’ve been testing it in early access across complex coding, debugging, and reasoning workflows. On selected internal evals, we saw encouraging improvements: - 12% higher task completion rate - 16% fewer retries on complex…

Tip 3: There is a big jump on coding tasks like - SWE-Bench Pro 64.3% - SWE-Bench Verified 87.6% - TerminalBench 69.4% Altogether, it’s a strong signal that coding performance has improved meaningfully in Opus 4.7 compared to Opus 4.6.

Tip 2: Opus 4.7 shows improvement on languages including those with less training data like Yoruba, Igbo and Chichewa. For folks who speak these and other languages this model will be meaningfully better.

Tip 1: Opus 4.7 uses a new tokenizer behind the scenes. In our testing, it can use between 1x and 1.5x more tokens when compared to Opus 4.6, so it’s worth planning accordingly.