Atlas

Benchmarks

← All benchmarks

Clipping Bench v1.0

Games

Naïve's simulated month running a short-form clip farm. From $300 and three social accounts, a model joins brand campaigns, cuts clips from long videos and is paid for verified views while rival clippers drain the same budgets, over a season of 27 to 33 days that ends without warning. The score is settled cash, averaged over three seeds. Naïve's scripted reference strategy ends at $2,606, an oracle at $27,521, and doing nothing at $259.

Benchmark source: Usenaive

Top models (higher is better)

ModelScoreSource
GPT-6 Astra9514Clipping Bench v1.0 leaderboard: GPT-6 Astra · Naïve
GPT-6.1 Sol5707Clipping Bench v1.0 leaderboard: GPT-6.1 Sol · Naïve
Claude Opus 5.52018Clipping Bench v1.0 leaderboard: Claude Opus 5.5 · Naïve
GPT-6 Sol1240Clipping Bench v1.0 leaderboard: GPT-6 Sol · Naïve
Claude Fable 5.11068Clipping Bench v1.0 leaderboard: Claude Fable 5.1 · Naïve
Claude Sonnet 5.5886Clipping Bench v1.0 leaderboard: Claude Sonnet 5.5 · Naïve
Grok 4.6805Clipping Bench v1.0 leaderboard: Grok 4.6 · Naïve
Claude Opus 5734Clipping Bench v1.0 leaderboard: Claude Opus 5 · Naïve
GLM-5.3644Clipping Bench v1.0 leaderboard: GLM 5.3 · Naïve
Gemini 3.8 Flash596Clipping Bench v1.0 leaderboard: Gemini 3.8 Flash · Naïve
Qwen3.6 Plus (2026-04-02)458Clipping Bench v1.0 leaderboard: Qwen 3.6 Plus · Naïve
GLM-5.3 Flash399Clipping Bench v1.0 leaderboard: glm-5.3-flash · Naïve
Grok 4.7384Clipping Bench v1.0 leaderboard: Grok 4.7 · Naïve
GPT-6 Luna374Clipping Bench v1.0 leaderboard: GPT-6 Luna · Naïve
Kimi K3343Clipping Bench v1.0 leaderboard: Kimi K3 · Naïve
MiniMax M3223Clipping Bench v1.0 leaderboard: MiniMax-M3 · Naïve
Gemini 3.5 Flash218Clipping Bench v1.0 leaderboard: Gemini 3.5 Flash · Naïve
Loading Atlas data…