Mixed results against Opus 5

ALL-AI reports that Moonshot AI’s Kimi K3.1 Pro scored 70% on Humanity’s Last Exam with tools, compared with 68% for Opus 5.5 and 57% for GPT-6 Astra. The ordering changes elsewhere: Kimi’s 1,846 Elo in GDPval-AA v2.1 trails Opus at 1,866. Its 66% Terminal-Bench 4.0 result also falls just below Opus.

API access and a planned weights release

The report describes a one-million-token context window and a focus on extended agent workflows. It lists standard API prices of $2 per million input tokens and $10 per million output tokens. API access is already available, while the full model weights are expected by November 4, 2026. That planned release does not demonstrate a further improvement in model performance.

What the comparison does not establish

These are results for particular evaluations, not proof that Kimi is universally the best model or that employees will become more productive. The report also notes that an independent Artificial Analysis assessment did not place Kimi first. Readers should distinguish the provider’s reported results from independently reproduced measurements and from outcomes in their own tasks.

Limits of the evidence

This review summarizes the supplied report; it does not independently reproduce the benchmarks or measure workplace outcomes.

Source

https://www.all-ai.de/news/news26top/kimi-k3-1-pro-ki-modell-start