Kimi K2.5 is the first leading open-weights model with image input and ranks just behind OpenAI and Anthropic on GDPval-AA, at roughly a quarter of Opus 4.5's evaluation cost.
Key takeaways · AI-distilled
Its GDPval-AA Elo of 1309 implies a 66% win rate against GLM-4.7, the prior open-weights leader, on knowledge-work tasks run with shell and web access through Artificial Analysis's Stirrup agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition →.
K2.5 scores 75% on the MMMU Pro visual reasoning benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition →, slightly behind Gemini 3 Pro and in line with GPT-5.2 and Claude Opus 4.5, while open rivals DeepSeek V3.2, GLM-4.7, MiniMax M2.1 and MiMo-V2-Flash lack image input.
It merges Moonshot's reasoning and non-reasoning lines into one hybrid model; Artificial Analysis benchmarked it with reasoning on and said reasoning-off results would follow.
Its hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition → rate fell to 64% from K2 Thinking's 74%, which drives the -11 AA-Omniscience score: the model abstains slightly more often instead of fabricating answers when uncertain.
It used about 82M reasoning tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → on the Intelligence Index, below K2 Thinking (~95M) and GLM-4.7 (~160M). Its $371 run cost is over 4x cheaper than Opus 4.5 and GPT-5.2 but over 5x DeepSeek V3.2 and gpt-oss-120b.
Terms in this piece · Glossary
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
