Benchmarks
The models we run, measured.
We pick models for results, then make them affordable. Here's how the models available on Vikasit perform against the global frontier — competitive on math, honest about the gap on the very newest closed models, and far ahead on cost.
Running with the frontier
On AIME 2025, the flagship we serve is within ~3 points of the best — and far ahead on cost.
No tools. The newest Anthropic/Google flagships no longer report AIME — shown are the most recent that do.
Gemini 3 Pro
GPT-5.1
vikasit-235b-moe
MiniMax-M2AIME'25 from vikasit-30b-moe — just ~3B active params.
MATH from vikasit-math (72B). 7B hits 85.3%.
A tier for every budget.
From compact models to a 235B-class flagship — and the trillion-scale tiers above them — quality climbs cleanly with scale. Our efficient MoE options deliver frontier-class results at small-model inference cost, so you only pay for what you need.
Thinking mode. The 30B MoE (3B active) rivals far bigger dense models.
vikasit-235b-moevikasit-30b-moevikasit-32bvikasit-14bvikasit-8bThe bar keeps rising
The newest closed models (June 2026) have pushed coding and graduate reasoning higher. Here's where they stand — and where the best models on Vikasit sit today. We show it straight.
Latest closed frontier vs the best model on Vikasit. No tools.
Gemini 3.1 Pro
Claude Fable 5Claude Opus 4.8vikasit-235b-moeOur coding-tuned tier targets the open state of the art.
Claude Mythos 5Claude Fable 5Claude Opus 4.8Gemini 3.1 Provikasit-titan-1.6t