Qwen 3 Max Thinking

Model details

Model: alibaba/qwen3-max-thinking
Provider: vercel-ai-gateway
API: anthropic-messages
Base URL: https://ai-gateway.vercel.sh
Input: text
Reasoning: Yes
Context window: 256,000
Max tokens: 65,536

Show configuration

{
  "providers": {
    "vercel-ai-gateway": {
      "apiKey": "YOUR_API_KEY",
      "models": [
        {
          "id": "alibaba/qwen3-max-thinking",
          "name": "Qwen 3 Max Thinking",
          "reasoning": true,
          "input": [
            "text"
          ],
          "contextWindow": 256000,
          "maxTokens": 65536,
          "cost": {
            "input": 1.2,
            "output": 6,
            "cacheRead": 0.24,
            "cacheWrite": 0,
            "tiers": [
              {
                "inputTokensAbove": 32000,
                "input": 2.4,
                "output": 12,
                "cacheRead": 0.48,
                "cacheWrite": 0
              },
              {
                "inputTokensAbove": 128000,
                "input": 3,
                "output": 15,
                "cacheRead": 0.6,
                "cacheWrite": 0
              }
            ]
          },
          "compat": {
            "allowEmptySignature": true
          }
        }
      ],
      "api": "anthropic-messages",
      "baseUrl": "https://ai-gateway.vercel.sh"
    }
  }
}

Pricing

USD per million tokens. A tier is selected from the total input tokens in each request and applies to that entire request.

Pricing rates for Qwen 3 Max Thinking
Request input	Input	Output	Cache read	Cache write
Up to 32,000 tokens	$1.2	$6	$0.24	$0
Above 32,000 tokens	$2.4	$12	$0.48	$0
Above 128,000 tokens	$3	$15	$0.6	$0

Session cost calculator

Estimate the requests made during an agent session. A user turn can make several model calls while using tools, so costs are calculated per model request.

Session shape

Rounded median shapes from recent Pi agent sessions.

Model requestsStarting contexttokensContext added / requesttokensOutput / requesttokens

Warm-request cache reuse

96%Effective token reuse combines partial hits and occasional complete misses.

Starting context is already cached

Estimated session$0.00

Uncached input: —
Cache reads: —
Uncached prefixes: —
Cache writes: —
Output: —
Without caching: —

The first request starts cold unless marked otherwise. Output is billed separately from context growth because reasoning tokens are not always retained. This remains a directional estimate: providers differ in eligibility, rounding, and retention.

Compatibility flags

Effective values after applying Pi's API defaults and model overrides.

Effective compatibility flags for Qwen 3 Max Thinking
Feature	Value
`supportsEagerToolInputStreaming`	Yes
`supportsLongCacheRetention`	Yes
`sendSessionAffinityHeaders`	No
`supportsCacheControlOnTools`	Yes
`supportsTemperature`	Yes
`forceAdaptiveThinking`	No
`allowEmptySignature`	Yes

Also available from other providers