ScreamingFace — The Forge (direction 2)

the proof · latest knock-down

Humanity's Last Exam · multiple-choice · 513 questions · Mar 23, 2026

ScreamingFace Fusion 😱50.3%

Gemini 3.1 Pro44.4%

Claude Opus 4.639.8%

Gemini 3.0 Pro36.6%

GPT-5 Pro30.2%

Kimi K2.529.6%

✓ every score enclave-verified · fusion +5.8 over the best single model · verify yourself ↗

never run out of tokens again

Point your tools at the model that wins.

Aim Claude Code, Gemini CLI, or your own app at ScreamingFace and every prompt routes to the model most likely to win it — so your tokens stretch further, and when you run dry you fall back to open models instead of stopping. Download for the axis you care about; the first run tells you what you just saved.

accuracy

Always route to the model most likely to be right.

“✓ beat SOTA by +5.8”

download ↓ · proven today

cost

Spend the cheapest model that still clears the bar.

“you saved 60% of your spend”

download ↓ · on the roadmap

speed

Route to the fastest model that holds quality.

“you just saved 5 seconds”

download ↓ · on the roadmap

curl screamingface.ai/install | shone line, in the tool you already use.

what's actually running

One model, made of models.

A fusion model is one model, made of models. Each one votes; we weight every vote by how it scored on a handful of calibration questions, then aggregate with differential-privacy noise so no single source can dominate the answer. A mixture of experts — made honest. And the whole thing is a URL you can run.

the fusion model · as a URLurl4

# the data
data.url4.ai/hle.jsonl
  / google/gemini-3.1-pro      ! $q :vote<A|B|C|D|E>
  | anthropic/claude-opus-4-6  ! $q :vote<A|B|C|D|E>
  | google/gemini-3-pro        ! $q :vote<A|B|C|D|E>
  | openai/gpt-5-pro           ! $q :vote<A|B|C|D|E>
  | moonshotai/kimi-k2.5       ! $q :vote<A|B|C|D|E>
  / enclave.url4.ai/gnmax?σ=5.0  ! $vote

✓ enclave-verifiedε ≤ 40.6 per sourcere-runnable, forkable, PR-able

the scream team

Help build the Wikipedia of taste.

We knock down a new benchmark every week. Benchmarks are community-contributed — every one is a definition of good, and the world needs millions of them. Bring one, beat one, or just get each verified result the day it lands.