New Sign up free, 10 calls on us. Up to $1, no card needed.

Qwen3.7 Max

Released 2026-05-21

chatCodeReasoningTool callingPrompt caching

Qwen3.7 Max is the flagship of the Qwen team's "Agent Frontier" generation, described by Alibaba as its most advanced and comprehensive agent model to date.

Input
text $2.5/M
Output
text $7.5/M
Cache read
$0.5/M
Context
1M
vs GPT-4o
~50% cheaper

Benchmarks

Above averageNo peer higher43 / 9112 / 91
Qwen3.7 Max other models measured peer average no peer scored higher
SWE-Bench Pro
60.6%
AndroidBench
56.5%
HealthBench
54.5%
JobBench
31.3%
PLawBench
58.9%
GPQA Diamond
92.4%
MCP-Atlas
76.4%

Vendor-published: Alibaba (Qwen) Anthropic ByteDance DeepSeek Google MiniMax Moonshot OpenAI Tencent Z.ai

Price in context

Where the price sits among 60 comparable models

Input$2.5/M
$0.05 · Qwen3 VL Flash GPT-5.4 Pro · $30
Output$7.5/M
$0.275 · DeepSeek V4 Flash GPT-5.4 Pro · $180
Cached read$0.5/M
$0.0028 · DeepSeek V4 Flash GPT-5.4 Pro · $15

The bar shows how this model’s price compares with every other model of the same kind on Synthorai. The cheapest and the most expensive are named at each end. These are base rates; batch, region and cache-write discounts are on the pricing page.

Specs & limits

Tokens

Context window (vendor spec) 1,000,000
Max output (vendor spec) 65,536

Prompt caching

How it caches automatic + explicit
Min prefix 1,024
Lifetime explicit: 5m, reset on hit
Write cost 1.25x

Thinking

Vendor control enable_thinking + thinking_budget + preserve_thinking
Accepted values enable_thinking true · false; thinking_budget in tokens; preserve_thinking true · false
Default on; the Qwen3.7 Max series is hybrid thinking with thinking enabled by default, and preserve_thinking is off applied when the request sets nothing
Can be turned off Yes
Thinking behaviour Trace returns in reasoning_content and is dropped from later turns unless preserve_thinking is true; the qwen3.7-max-preview snapshots are thinking-only and cannot disable it.
Parameter reasoning_effort
Values minimal · low · medium · high the gateway's parameter surface - the vendor mapping above applies

Model

Modalities text → text
  • "Agent Frontier" flagship: text-only, thinking on by default, 1M-token context
  • built for long-horizon autonomous agent runs (officially demonstrated 35-hour session with 1,158 tool calls)
  • not listed for JSON mode

per Alibaba official docs ↗

One prompt, measured through the gateway

PROMPT Explain to a CFO, in exactly three sentences, why our LLM bill doubled after we enabled reasoning. CHECK exactly 3 sentences

Qwen3.7 Max passed · 3 sentences

Enabling reasoning requires the model to generate extensive, step-by-step internal logic before delivering a final answer, drastically increasing our consumption of premium-priced output tokens. Because AI providers charge significantly higher rates for generated text than for input prompts, this surge in behind-the-scenes computation directly multiplied our baseline API costs. However, this increased spend delivers a strong ROI by successfully automating complex workflows that previously required expensive human intervention.

out 1700 tok (+1613 thinking) latency 28.9 s

Instruction following (exactly three sentences - countable), audience shaping (CFO register), and the hidden-thinking billing gap the token meters below expose.

PROMPT This function is supposed to return the longest run of consecutive equal items, but callers report it is off by one on single-element inputs and misses runs that end at the last index. Fix it and explain the bug in one paragraph. CHECK fix passes tests

Qwen3.7 Max passed · 8/8 cases

## Fixed Function ```python def longest_run(xs): if not xs: return 0 best = 1 cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 best = max(best, cur) else: cur = 1 return best ``` ## Explanation The bug is that `cur` and `best` are initialized to `0`, which makes the function count adjacent equal **pairs** (transitions) rather than the number of **elements** in a run. A run of *n* consecutive equal items contains only *n − 1* adjacent pairs, so the reported length is always one less than the true length. This causes single-element inputs (and any list with no adjacent duplicates) to return `0` instead of `1`, and causes every longer run—including those that extend to the final index—to be undercounted by one. […]

out 3433 tok (+3149 thinking) latency 50.6 s

Whether the fix is actually correct (runnable), explanation density, and token efficiency on a bounded task.

PROMPT Extract a JSON object with fields {company, quarter, revenue_usd_m, growth_pct, guidance} from this text. Use null for anything not stated; add no other fields. CHECK valid JSON, schema exact

Qwen3.7 Max passed · 5/5 fields, guidance "withheld"

{ "company": "Meridian Robotics", "quarter": "September quarter", "revenue_usd_m": 412, "growth_pct": 11, "guidance": "withheld" }

out 1595 tok (+1534 thinking) latency 26.9 s

Schema obedience (no invented fields), hallucination pressure (guidance is explicitly withheld), and structured-output path differences.

PROMPT Write a 120-word product announcement for an API gateway feature that caches prompts across providers. Forbidden words: "seamless", "unlock", "game-changer", "revolutionize", "empower". Exactly one sentence must be a question. CHECK 120 words, 0 banned words

Qwen3.7 Max passed · 120 words, 0 banned, 1 question

Introducing Global Prompt Caching for our API gateway. Why pay multiple artificial intelligence vendors for the exact same query? Our new feature intercepts identical requests before they reach external models, serving cached responses instantly. This reduces your monthly bills significantly while dropping latency to near absolute zero. Developers routing traffic between various large language models now share a unified memory layer. When a user asks a repeated question, the gateway retrieves the stored answer, bypassing redundant compute entirely. You maintain complete control over cache expiration and invalidation rules directly through your main configuration dashboard. Stop wasting your expensive tokens on duplicate prompts today. Upgrade your infrastructure right now and watch your operational costs drop while your overall application speed increases.

out 8147 tok (+8005 thinking) latency 114.3 s

Constraint obedience (word budget, banned-word list, the single question), style fingerprint, and length control.

Use Qwen3.7 Max in 30 seconds

OpenAI-compatible: swap the base_url, keep your SDK. POST /v1/chat/completions

from openai import OpenAI

client = OpenAI(
    base_url="https://synthorai.io/v1",
    api_key="sk-syn-...",
)

resp = client.chat.completions.create(
    model="qwen3.7-max",
    messages=[{"role": "user", "content": "Summarize this diff"}],
    reasoning_effort="medium",
)
print(resp.choices[0].message.content)

About Qwen3.7 Max

  • It brings a 1M-token context window, double the previous generation's preview limit, and is built for long-horizon autonomous execution spanning hundreds or even thousands of steps, including coding and debugging and office workflow automation.
  • Alibaba's own demonstration of that claim is a single session running about 35 hours across more than a thousand tool calls.
  • It focuses on text input with deep reasoning and tool invocation (the Plus tier is where vision lives) and returns up to 65,536 output tokens, on top of the largest documented reasoning allowance in the Qwen line, doubled again over Qwen3.6.
  • Alibaba's selection guidance is blunt about when to spend on it: choose Max when you need the strongest reasoning, and Plus otherwise.
  • Thinking is hybrid and enabled by default, switched off with enable_thinking and bounded with thinking_budget, with the trace returned in reasoning_content and billed as output; note that the preview snapshots of this model are thinking-only and cannot turn it off.
  • The generation's distinctive control is preserve_thinking, which carries reasoning across turns so an agent's earlier deliberation stays available instead of being re-derived, and Alibaba names this model among the handful that support it, which matters more here than anywhere else given the session lengths it targets.
  • Function calling, batch inference and both explicit and implicit caching are supported; JSON mode is not listed.
  • Synthorai puts Qwen3.7 Max behind its OpenAI-compatible API for seamless adoption.

FAQ

Is the Qwen3.7 Max API free to try?

Yes: new accounts get 10 trial calls and up to $1 in free credit, no card required. At $2.5/M input tokens, that credit alone covers roughly 49 requests of ~8K tokens against Qwen3.7 Max.

What is Qwen3.7 Max best at?

1M context, double the previous preview limit; long-horizon execution spanning thousands of steps; most advanced and comprehensive agent model yet. See the About section for the full picture from the vendor's own release notes.

How much does Qwen3.7 Max cost?

Qwen3.7 Max costs $2.5 per million input tokens and $7.5 per million output tokens on Synthorai. That is the provider's list price, with no platform markup. Cached input tokens bill at $0.5/M.

Does Qwen3.7 Max support prompt caching?

Yes: automatic caching is on by default, with an explicit mode for guaranteed savings. Cached input tokens bill at $0.5/M vs $2.5/M uncached; prompts need a 1,024-token stable prefix to cache (TTL explicit: 5m, reset on hit). Prompt caching guide →

How do I get access to Qwen3.7 Max?

Point your existing OpenAI SDK at base_url="https://synthorai.io/v1", set model="qwen3.7-max", and you're done. One API key covers every model on the gateway.

Related models

Compare

Every value on this page is transcribed from the vendor's own documentation, linked above, and carries the date it was checked. Prices are compared across the catalogue; specification values that vendors define differently are shown with the difference stated rather than charted. Nothing here is measured by us, and nothing is scored.

Get your API key Compare your cost →