Step 5 Preview on QuickSilver Pro
Step 5 Preview brings reasoning, tools and image understanding to a 1M-token context. OpenRouter list parity: $1.00 input / $2.70 output per million tokens, with $0.10 cached input.
At a glance
Reasoning over long documents and images, tool workflows and structured output.
Pricing comparison ($/1M tokens)
| Provider | Input | Output | vs QSP |
|---|---|---|---|
| QuickSilver Pro | $1.00 | $2.70 | — |
| OpenRouter list price (stepfun/step-5-preview) | $1.00 | $2.70 | same |
When to use
Use for text and image understanding, multi-step tool workflows and strict JSON Schema output. Streaming, automatic tool calls and forced named functions were verified on the pinned endpoint.
When to use something else
Choose another model when you need reasoning disabled. This preview requires reasoning; small output budgets can be consumed entirely by reasoning before a visible answer appears. Video is listed upstream but has not been validated on this route.
Quickstart (curl)
curl https://api.quicksilverpro.io/v1/chat/completions \
-H "Authorization: Bearer $QSP_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "step-5-preview",
"messages": [{"role": "user", "content": "Hello!"}]
}'OpenAI-compatible. One-line migration via base_url.
FAQ
No. The live probe returned reasoning by default and rejected reasoning: {enabled: false}. reasoning: {enabled: true} and reasoning: {effort: "low"} worked through OpenRouter. The output budget includes reasoning; max_tokens=16 and 64 were both exhausted by reasoning in a harder prompt.
$1.00 input / $2.70 output per million tokens, with $0.10 per million cached input tokens. These are OpenRouter list prices, with discount factor 1.0. The pinned provider currently has a 5% promotion; that temporary provider discount does not change the list-parity selling rule.
The pinned endpoint lists 1M context tokens and up to 66K output tokens. Text, image input, streaming, tools, forced named functions and strict response_format json_schema were verified. The advertised limits come from endpoint metadata; the probe used at most 256 output tokens.
Use base_url=https://api.quicksilverpro.io/v1 and model="step-5-preview" with the OpenAI SDK. Chat Completions and Responses share this route. Requests are pinned to DeepInfra FP8 with zero data retention and no provider fallbacks. The official StepFun endpoint failed the ZDR filter and does not support forced tool choice.