Mean absolute gap in percentile positions on shared benchmark settings; lower is closer. At least 60% overlap is required, with a minimum of three when available. Independent evaluations are used when there are at least three; otherwise published report comparisons are used. Repeated model observations within one setting are averaged before ranking. Writing-style traits and composite indices are excluded. This describes similarity, not an overall quality ranking.

About the model report

Earlier-2026 baseline cohort from the official model card. Asterisks identify competitor scores re-evaluated by Kimi rather than cited from their reports. Coding scores average ten runs; vision averages three runs. Tool-enabled and no-tools results, F1 and exact accuracy, pass@3 and pass^3 remain separate rows. MCPMark here predates the Verified default. APEX uses 452 public tasks. Reviewed 2026-10-26.

Read the official model report ↗

25 benchmarks in the report

Search this collection →

SCI / BENCHMARK258 reported results

SciCode

Scientist-authored coding problems require translating domain knowledge into numerical implementations.

Executable correctness↗

RSN / BENCHMARK131 reported results

Humanity’s Last Exam

Expert-authored questions stretch academic reasoning and knowledge across many disciplines.

Answer accuracy↗

SCI / BENCHMARK507 reported results

GPQA Diamond

A demanding subset of expert-written science questions tests detailed reasoning in specialist domains.

Multiple-choice accuracy↗

RSN / BENCHMARK146 reported results

MMMU-Pro

Academic questions require visual understanding as well as language reasoning, with measures against text-only shortcuts.

Multimodal accuracy↗

COD / BENCHMARK180 reported results

LiveCodeBench

Date-bounded contest problems test coding and related code-repair or execution skills.

Pass@1↗

COD / BENCHMARK133 reported results

SWE-bench Verified

Human-reviewed repository issues test whether an agent can implement a repair in an existing Python codebase.

Resolved issues↗

COD / BENCHMARK5 reported results

SWE-bench Pro

Long-horizon software-engineering tasks extend repository evaluation beyond small isolated edits.

Resolved issues↗

AGT / BENCHMARK29 reported results

OSWorld-Verified

Language-model agents act in real computer environments across applications, using reproducible setup and evaluation scripts.

Verified task success↗

AGT / BENCHMARK12 reported results

BrowseComp

Difficult information-seeking questions test a browsing agent’s ability to locate and combine evidence.

Answer accuracy↗

RSN / BENCHMARK20 reported results

BabyVision

Visual problem solving tests whether a model can reason over image structure rather than rely only on familiar textual knowledge.

Visual reasoning accuracy↗

AGT / BENCHMARK10 reported results

DeepSearchQA

Web research tasks require collecting a complete set of answers rather than finding a single plausible fact.

Answer-set F1 and exact accuracy↗

AGT / BENCHMARK8 reported results

WideSearch

Broad information gathering asks an agent to identify many entities and their attributes across web sources.

Item-level F1↗

AGT / BENCHMARK5 reported results

Toolathlon

Long-horizon tasks require coordinating tools and completing a practical workflow.

End-to-end tool task success↗

AGT / BENCHMARK11 reported results

MCPMark

Agents work through MCP-connected application and filesystem tasks with programmatic verification.

Verified tool task success↗

AGT / BENCHMARK10 reported results

Claw Eval

Tool-based agent evaluations report both repeatability and success across multiple attempts.

Pass^3 and pass@3↗

COD / BENCHMARK280 reported results

Terminal-Bench 2.0

An agent completes technical work in a terminal under executable task checks.

Verified terminal task success↗

COD / BENCHMARK10 reported results

SWE-bench Multilingual

Repository-level software repair broadens evaluation across programming languages.

Resolved repository issues↗

COD / BENCHMARK4 reported results

OJBench

Programming problems are solved against an online-judge-style evaluation, including a named Python setting in model reports.

Online-judge programming accuracy↗

MTH / BENCHMARK11 reported results

AIME 2026

A new annual contest set tests mathematical reasoning through short integer answers.

Contest answer accuracy↗

MTH / BENCHMARK5 reported results

HMMT 2026 · February

The February Harvard–MIT Mathematics Tournament provides a dated contest-math evaluation.

Contest answer accuracy↗

MTH / BENCHMARK11 reported results

IMO-AnswerBench

Olympiad-related mathematics is evaluated through answer-oriented scoring in frontier model reports.

Answer accuracy↗

RSN / BENCHMARK26 reported results

CharXiv

Scientific charts require models to read visual information and reason over what the figure actually shows.

Chart reasoning accuracy↗

MTH / BENCHMARK16 reported results

MathVision

Mathematical problems with visual context require reasoning over both the statement and the supplied figures.

Visual mathematics accuracy↗

RSN / BENCHMARK5 reported results

V*

Detailed visual questions test the ability to locate and interpret relevant evidence inside an image.

Visual search accuracy↗

AGT / BENCHMARK66 reported results

APEX-Agents · original

Professional-service workflows require sustained work across files and tools.

Professional-services task success↗

COMPARISON CHECKLIST

Match benchmark version, task subset, model effort, harness, tool access, attempt budget, scoring rule, and fallback behavior. A shared benchmark name is the start of a comparison, not proof that two reported scores use the same contract.

Kimi original report ↗