Qwen3.8 Max on QuickSilver Pro
Qwen 3.8 Max is Alibaba's August 2026 flagship and the largest model it has shipped — 2.4 trillion parameters in a mixture-of-experts, with a 1M-token context window and text, image and video input. Alibaba positions it for autonomous coding and long-horizon agentic work rather than chat. On QuickSilver Pro it's $2.00 input / $6.00 output per 1M tokens — Alibaba's own published rate, passed through without a markup — with one OpenAI-compatible key across the whole catalog.
At a glance
Autonomous coding + long-horizon agents — Alibaba's largest model, 1M context, at its own list price.
Pricing comparison ($/1M tokens)
| Provider | Input | Output | vs QSP |
|---|---|---|---|
| QuickSilver Pro | $2.00 | $6.00 | — |
| Alibaba (qwen3.8-max) | $2.00 | $6.00 | same |
| OpenAI (GPT-5.6 Sol) | $5.00 | $30.00 | 80% lower |
When to use
Reach for 3.8 Max on work that runs unattended for a long time: multi-day coding agents, planning loops that span hundreds of turns, and tasks where the agent has to notice its own mistakes and correct course rather than being re-prompted. Alibaba's launch material describes multi-day autonomous development runs and planning loops in the hundreds of turns, and pitches the vision path as a feedback loop for planning and self-correction rather than input alone — treat those as vendor claims and match them against your own evals. The 1M context means a large repository or a long trace can stay resident instead of being re-summarised each turn.
When to use something else
This is the top of the range and priced like it. For routine chat, codegen or production agents, Qwen 3.7 Plus at $0.256/$1.024 is roughly 6× lower on output and Qwen 3.7 Max at $1.25/$3.75 about 1.6× lower — start there and step up only where your evals show 3.8 Max earning it. It also thinks by default: the QuickSilver Pro gateway sends enable_thinking=false so routine calls don't bill a hidden reasoning trace, which matters more here than on lower-cost models. Pass enable_thinking=true where you want the trace, and budget output tokens accordingly.
Quickstart (curl)
curl https://api.quicksilverpro.io/v1/chat/completions \
-H "Authorization: Bearer $QSP_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-max",
"messages": [{"role": "user", "content": "Hello!"}]
}'OpenAI-compatible. One-line migration via base_url.
FAQ
$2.00 per 1M input tokens and $6.00 per 1M output tokens, with cached input reads at $0.25 per 1M. That is Alibaba's own published rate, passed through unchanged. Most of our catalog sells below the market reference, but Alibaba is the only place this model exists, so there is no lower-cost route to fund a discount and we do not pretend otherwise — you pay the same as at the source, with one key and one balance across every other model we carry.
On list prices, yes, substantially: $2.00/$6.00 against OpenAI's $5.00/$30.00 for GPT-5.6 Sol — 80% less on output — and against Anthropic's $5.00/$2.005.00 for Claude Opus 5. Those are different models from different labs, so price alone should not decide it; benchmark all three on your own workload. If you want to compare on QuickSilver Pro directly, both GPT-5.6 Sol and Claude Opus 5 are on the same API key and the same balance, so switching is a one-line model-string change.
It thinks by default. Measured against the live endpoint on launch day, answering "Reply with exactly: OK" cost 28 completion tokens, 22 of them an invisible reasoning trace. To stop routine calls billing for that, the QuickSilver Pro gateway sends enable_thinking=false by default. Pass enable_thinking=true in the request body to opt back in — it works on non-streaming requests as well as streaming, which is not true of every Qwen model.
1M tokens of context. Alibaba documents a maximum input of 991K tokens, dropping to 983K when thinking is enabled, and a maximum output of 131K tokens. Cached input is billed at $0.25 per 1M rather than the full $2.00, so a long system prompt or a repository kept resident across turns costs far less on the second and subsequent calls than on the first.
3.8 Max is the newer and much larger model — 2.4T parameters against the 3.7 line's flagship — and Alibaba aims it specifically at autonomous coding and long-horizon agentic work, with vision as part of the planning loop rather than just an input format. Both expose 1M context. 3.7 Max is 1.6× lower on output ($1.25/$3.75) and remains the better default for hard reasoning that does not need to run unattended for days; step up to 3.8 Max where your evals show the longer-horizon behaviour paying for itself.
Not at launch. Alibaba announced that the weights for Qwen 3.8 Max, and for the smaller Qwen 3.8-27B, would follow within about a week of the 2026-08-03 release. Until then the hosted API is the only way to run it. We will keep serving it through the same endpoint either way — nothing about your integration changes if and when the weights land.
Point the OpenAI SDK at https://api.quicksilverpro.io/v1 and set the model string to qwen3.8-max. Streaming, tool / function calling and structured JSON output all work on the standard Chat Completions surface — tool calls were verified against production on launch day. Every response carries OpenAI-standard usage accounting plus a usage.cost field computed from the public prices above, so you can reconcile spend client-side without a second API call.