Introducing PaladinOur first field-based AI model

Closed general intelligence.

A frontier AI lab developing field-based intelligence.

Paladin, against the frontier

6 models · 5 benchmarks

#5Paladin trails by 4.7

Graduate-level science questions.

Paladin, against the frontier. Demo numbers, not real results.
Benchmark	Paladin	Claude	GPT	Gemini	Grok	DeepSeek
GPQA Diamond	78.4	80.2	83.1	81.0	79.3	74.2
SWE-bench Verified	74.6	77.0	72.4	66.1	63.0	64.3
AIME 2026	91.3	83.7	90.0	86.7	88.1	82.0
τ²-bench	79.8	75.1	70.4	64.9	60.2	58.3
RULER 1M	93.1	84.0	80.2	90.4	74.5	70.1

*Placeholder figures, not measured results.

Paladin, against a chat model.

Time per task

Each square is one Paladin task. The grid is one chat-model task.

Cost per task

Each square is one Paladin task. The grid is one chat-model task.

*Placeholder figures, not measured results.

Context window

Context windowHow much each model takes in at once, in tokens, on one scale. Paladin holds 10 million.

Hallucinations

HallucinationsHow often each model states something its inputs don’t support, on the same workflows. Paladin: none.