Qwen3.7 Max on QuickSilver Pro
Qwen 3.7 Max is Alibaba's May 2026 flagship — the top of the Qwen 3.7 line, tuned for hard multi-step reasoning, long-horizon agentic workflows, and the deep tool-calling tasks where you'd otherwise reach for GPT-4o or Claude Sonnet 4.5. On QuickSilver Pro it's $1.25 input / $3.75 output per 1M tokens — ~62% lower than GPT-4o on output, with 1M context (8× GPT-4o's) and one OpenAI-compatible key across 32 models. That is Alibaba's own price for the model, passed through.
At a glance
Flagship reasoning + long-horizon agentic — GPT-4o quality at ~62% lower output cost, 8× the context.
Pricing comparison ($/1M tokens)
| Provider | Input | Output | vs QSP |
|---|---|---|---|
| QuickSilver Pro | $1.25 | $3.75 | lowest-cost |
| OpenRouter (qwen/qwen3.7-max) | $1.475 | $4.425 | 15% lower |
| OpenAI (GPT-4o) | $2.50 | $10.00 | 63% lower |
When to use
Reach for 3.7 Max when 3.6 Plus isn't smart enough on your eval: multi-hop reasoning, agentic workflows that plan across 10+ tool calls, long-context analysis over a million-token corpus, code-review or refactor agents that need to hold a whole repo in head. Same shape as GPT-4o or Claude Sonnet 4.5 on the workload — pick 3.7 Max for the cost win on output tokens and the wider context window.
When to use something else
For routine chat, codegen, or production agents where DeepSeek V4 Pro at $0.70/$2.10 lands the same answer, V4 Pro is ~6× lower on output. For pure math / theorem-style reasoning, DeepSeek V4 Pro is the price/quality leader. 3.7 Max thinks by default; the QuickSilver Pro gateway turns thinking off by default so routine calls don't bill a hidden reasoning trace — pass `enable_thinking=true` to opt back in.
Quickstart (curl)
curl https://api.quicksilverpro.io/v1/chat/completions \
-H "Authorization: Bearer $QSP_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.7-max",
"messages": [{"role": "user", "content": "Hello!"}]
}'OpenAI-compatible. One-line migration via base_url.
FAQ
On published Qwen benchmarks (MMLU-Pro, GPQA Diamond, HumanEval, Tau-bench, BFCL-v3) 3.7 Max lands within striking distance of GPT-4o on reasoning and tool-use, and noticeably ahead on long-context retrieval thanks to its 1M-token window vs GPT-4o's 128K. Match on your own evals before committing — for the workloads 3.7 Max is built for (agentic + long context), the per-token economics ($1.25/$3.75 vs GPT-4o's $2.5/$10) and 8× context advantage are usually decisive.
3.7 Max thinks by default. To keep routine calls from billing a hidden reasoning trace, the QuickSilver Pro gateway turns thinking off by default on 3.7 Max requests. If you want the reasoning trace, pass `enable_thinking=true` in the request body and budget output tokens accordingly (3-5× typical chat output).
V4 Pro at $0.70 / $2.10 per 1M is ~6× lower on output and matches 3.7 Max on many published reasoning benchmarks (MATH, GPQA, AIME). Reach for 3.7 Max only when V4 Pro visibly underperforms on your agentic / long-context evals — Qwen's tool-use and BFCL-v3 scores edge ahead on the agent benchmarks Alibaba published. A/B before committing for the price premium; for most production reasoning workloads V4 Pro is the better default.