What the five Kayts models score on published benchmarks

Kayts gives you five models in one chat, from Kayts Mihr, fast and free, to Kayts Monte, our most powerful. Each one runs on a model built by another company, chosen by us for that job. Those makers test their models on public benchmarks and publish the scores. We copy those scores as they were printed; we did not run the tests ourselves.

One test, three models

Only three of the five makers report the same test, GPQA Diamond: 198 hard multiple-choice questions in biology, chemistry and physics, written so that experts in the field get most right and skilled non-experts with web access do not. It is one of the few places where the scores line up directly.

GPQA Diamond: science questions at PhD levelShare of questions answered correctly, as each maker published it. The three models whose makers report this test.

Kayts Mihr 81.7%

Kayts Anahit 89.2%

Kayts Nareia 90.1%

Model cards of the models behind Kayts Mihr, Anahit and Nareia, checked 20 October 2026.

Show the data as a table

	Value
Kayts Mihr	81.7%
Kayts Anahit	89.2%
Kayts Nareia	90.1%

All three answer most of these questions correctly; the spread is 8.4 points. Kayts Mihr, the one in every plan, sits closest to the bottom, and Kayts Nareia, from the Pro plan, at the top.

Every model, on the tests its maker chose

Most makers pick the tests that fit what their model is for: coding and terminal work for one, long documents for another. So each panel below has its own set, and a bar in one panel should not be read against a bar in another, even when the names look alike. Versions matter too: Terminal-Bench 2.1 and 4.0 are different tests, and Humanity's Last Exam with tools is far easier than without.

Published scores of each Kayts modelShare of each test passed, from 0 to 100 %, as the maker of the model behind it published it. Compare bars only within a panel.

Kayts Mihr

GPQA DiamondScience questions at PhD level 81.7%

MMLU-ProKnowledge across 14 fields 82.5%

HMMT February 2025Competition math 83.2%

LiveCodeBench v6Coding problems 65.6%

τ²-BenchUsing tools in a conversation 79.1%

MMMUQuestions about images, college level 78.4%

Kayts Anahit

GPQA DiamondScience questions at PhD level 89.2%

Humanity's Last ExamExpert questions, without tools 30.8%

LiveCodeBench v6Coding problems 90.3%

SWE-bench ProFixing real bugs in software projects 61.7%

Terminal-Bench 2.1Working in a terminal 73%

OSWorld-VerifiedWorking a computer by screenshots 84.3%

Kayts Siran

Terminal-Bench 4.0Working in a terminal, harder tasks 42.3%

DeepSWE v1.1Long software tasks in real open-source projects 77%

AutomationBench v1.0.6Business automations, run by Zapier 60.5%

Humanity's Last ExamExpert questions, without tools 42.6%

Agents' Last ExamLong agent tasks across many steps and tools 35.4%

Chartography, with toolsReading and making charts, with tools 83%

Kayts Nareia

GPQA DiamondScience questions at PhD level 90.1%

Humanity's Last ExamExpert questions, without tools 37.7%

LiveCodeBenchCoding problems 93.5%

SWE-bench VerifiedFixing real bugs in open-source projects 80.6%

BrowseCompFinding hard-to-find facts on the web 83.4%

MRCR 1MRecalling details across a million tokens 83.5%

Kayts Monte

Terminal-Bench 4.0Working in a terminal, harder tasks 66.4%

FrontierCode v1.1Patches to real open-source issues, graded for merge quality 54.4%

CursorBench 4.0Coding tasks from the Cursor editor 57.8%

Humanity's Last Exam, with toolsExpert questions, with web and code tools 67.7%

Terminal-Bench-Science 0.1Scientific work in a terminal 58.7%

OSWorld 2.1Working a computer by screenshots 81.8%

Sources: the model cards and announcements of the models behind each Kayts model, checked 20–21 October 2026 (listed in the Privacy Policy).