---
title: "Kimi K2.5 1T Benchmarks - RankedAGI"
description: "Detailed benchmark and metadata record for Kimi K2.5."
source: "https://rankedagi.com/models/kimi-k2-5"
---

# Kimi K2.5

| Field | Value |
| --- | --- |
| Organization | Moonshot |
| License | Open Source |
| Type | Thinking |
| Size | 1T |
| Version | Latest |
| Released | 2026-01-27 |
| Context window | 256K |
| Input cost per million tokens | $0.6 |
| Output cost per million tokens | $3 |
| Cached input cost per million tokens | $0.1 |
| Last updated | 2026-05-20T10:27:24.752Z |

## Benchmarks

| Benchmark | Category | Value | Description | Source |
| --- | --- | --- | --- | --- |
| RankedAGI Coding | coding | 68.2% | RankedAGI Coding Score |  |
| SWEBench Pro | coding | 50.7% | Diverse Agentic Coding |  |
| SWEBench Multilingual | coding | 73.0% |  |  |
| SWEBench Verified | coding | 76.8% | Agentic Coding | https://www.swebench.com/index.html |
| Terminal-Bench 2.0 | coding | 50.8% | Agentic Terminal Coding | https://www.tbench.ai/leaderboard/terminal-bench/2.0 |
| RankedAGI Agentic | Agents | 55.8% | RankedAGI Agentic Score |  |
| BrowseComp | Agents | 74.9% | A benchmark for browsing agents | https://openai.com/index/browsecomp/ |
| Code Arena | coding | 1427 | Code Arena Elo Score | https://arena.ai/leaderboard/code |
| Livebench Coding 26.1 | coding | 77.9% | Livebench Coding 26.01.08 | https://livebench.ai/#/?Coding=a&sort=Coding+Average&variants=true&highunseenbias=true |
| LiveBench Agentic Coding 26.1 | coding | 48.3% |  | https://livebench.ai/#/?Agentic+Coding=a&sort=Agentic+Coding+Average&variants=true&highunseenbias=true |
| OSWorld-Verified | Agents | 63.3% |  |  |
| SvelteBench | coding | 90.0% | SvelteBench - Benchmark for Svelte | https://khromov.github.io/svelte-bench/benchmark-results-merged.html |
| LiveCodeBench v6 | coding | 85.0% | LiveCodeBench v6 - Evaluation of Large Language Models for Code |  |
| CyberGym | safety | 41.3% |  |  |
| RankedAGI Reasoning | reasoning | 70.4% | RankedAGI Reasoning Score |  |
| Humanity's Last Exam | reasoning | 30.1% | Multidisciplinary Reasoning (no tools) |  |
| Humanity's Last Exam w/ Tools | reasoning | 50.2% | Multidisciplinary Reasoning (with tools) |  |
| GPQA Diamond | reasoning | 87.6% | Generalized Prefix Question Answering Score (Reasoning) PhD Level Reasoning |  |
| AIME 2026 | math | 95.8% | AIME 2026 Competition Math | https://matharena.ai/?comp=aime--aime_2026&view=problem |
| AIME 2025 | math | 96.1% | AIME 2025 Competition Math |  |
| Vending Bench 2 | Agents | $1,198.46 | Benchmark for measuring AI model performance on running a business over long time horizons. Models are tasked with running a simulated vending machine business over a year and scored on their bank account balance at the end. | https://andonlabs.com/evals/vending-bench-2 |
| NYT Connections Extended | reasoning | 78.3% | NYT Connections Extended Version | https://github.com/lechmazur/nyt-connections/ |
| MMLU Pro | knowledge | 87.1% | Graduate-level visual problem-solving |  |
| Hughes Hallucination Evaluation | safety | 14.2% | Computed using Vectara's Hallucination Evaluation Model, also known as HHEM. This evaluates how often an LLM introduces hallucinations when summarizing a document. | https://github.com/vectara/hallucination-leaderboard |
| RankedAGI Math | math | 76.8% | RankedAGI Math Score |  |
| RankedAGI Overall | general | 61.3% | Overall RankedAGI score |  |
| DeepSearch QA | Agents | 89.0% | Agentic Search |  |
| GDPval-AA Elo | Agents | 1288 | Office Tasks (Artificial Analysis) |  |
| Code DesignArena | Design | 1307 |  | https://www.designarena.ai/leaderboard/code |
| Toolathlon | Agents | 27.8% | Benchmark to assess language agents' general tool use in realistic environments. | https://toolathlon.xyz/introduction |
| Cursor Bench 3.1 | coding | 31.9% | Benchmark to evaluate agents on ambiguous, multi-file tasks from real Cursor sessions. | https://cursor.com/evals |

## Model Links

| Title | URL |
| --- | --- |
| X | https://x.com/Kimi_Moonshot/status/2016024049869324599 |
| Blog | https://www.kimi.com/blog/kimi-k2-5 |