Compare

Speed, side by side.

Pick a task and watch every provider answer it in real time.

Speed race

Prompt

Measured lanes replay real single runs. Modelled lanes use each provider’s published first‑token time and output speed. Sources sit under the table.

Vision

See what it sees, as it sees it.

Watch it live on your camera, or race it on a photo against SwiftInference and GPT‑4o Vision.

Ready

Nothing is recorded or uploaded.

What it sees

Waiting for the camera.

–

ms / frame

–

frames / s

Device

runs on

Side by side

How the options stack up.

ProviderModelFirst tokenOutput speedSource
OgbonLlama190 ms190 tokens/sOur measurement
SwiftInferenceLlama 3.1 8B468–853 ms15–18 tokens/sOur measurement
OpenAIGPT‑4o1,082–2,315 ms–Our measurement
AnthropicClaude Sonnet 4No answer–Our measurement
Cerebrasgpt‑oss‑120B1,410 ms1,753 tokens/sArtificial Analysis
Together AIgpt‑oss‑120B1,899 ms564 tokens/sGeneral Compute, AA
Fireworks AIgpt‑oss‑120B3,190 ms690 tokens/sArtificial Analysis
SambaNovagpt‑oss‑120B4,100 ms693 tokens/sArtificial Analysis
Groqgpt‑oss‑120B4,870 ms478 tokens/sArtificial Analysis
Dreamscale LabsDreamZero 14B320 ms per step–Their published claim
  • Measured: our single runs in October 2026 of Ogbon, SwiftInference, OpenAI and Anthropic, which returned no answer.
  • Cerebras, Fireworks, SambaNova and Groq: Artificial Analysis on gpt‑oss‑120B. Their first token includes the model’s thinking time.
  • Together: output speed from Artificial Analysis, first token from General Compute’s benchmark.
  • Dreamscale Labs: their published latency for one DreamZero 14B robotics step, not a chat API.
  • Models differ and figures change often, so check each provider before deciding.

Race your own model on your own traffic.

Send us the model that hurts most today. We benchmark it against your current provider and show you the numbers.