Qwen3.7 Max: A $2.50 Flagship With a Verbosity Tax
Alibaba's Qwen3.7 Max, announced May 20, is a closed-weight, API-only flagship with a 1M-token context window and a list price of $2.50 per million input tokens and $7.50 per million output. That is half the input price of Claude Opus 4.7 or GPT-5.5 and a quarter of their output price, for a model that trades blows with Opus 4.6 on Alibaba's own coding benchmarks.
The catch is verbosity. In Artificial Analysis's evaluation, Qwen3.7 Max generated about 97M tokens where the median model generated 24M. A cheap per-token rate can still produce an expensive bill if the model writes four times as much. Qwen3.7 Max is cheap per token; whether it is cheap per task depends on your workload.
What Alibaba claims
Alibaba's launch numbers, mostly compared against Opus 4.6 in max mode:
| Benchmark | Qwen3.7 Max | Comparison |
|---|---|---|
| SWE-bench Verified | 80.4 | Opus 4.6 Max: 80.8 |
| SWE-bench Pro | 60.6 | Kimi K2.6 Thinking: 59.5 |
| Terminal-Bench 2.0 (Terminus) | 69.7 | DeepSeek V4 Pro Max: 67.9 |
| GPQA Diamond | 92.4 | Opus 4.6 Max: 91.3 |
| Humanity's Last Exam | 41.4 | Opus 4.6 Max: 40.0 |
Alibaba's most striking demo was a 35-hour autonomous kernel-optimization run: 1,158 tool calls and a 10x geometric-mean speedup over the Triton reference. Like every long-horizon demo, it shows what the model can do under favorable conditions, not what it does on average.
Note what the comparisons avoid. The reference point is Opus 4.6, not the current Opus 4.7 or GPT-5.5. On LMRank, Qwen3.7 Max scores 9.0, behind Opus 4.7 (9.6) and GPT-5.5 (9.4) and level with DeepSeek V4 Pro.
The price comparison that matters
- Claude Opus 4.7 (9.6): $5 / $25. Better model, 3.3x the output price.
- GPT-5.5 (9.4): $5 / $30. Better model, 4x the output price.
- Gemini 3.1 Pro Preview (8.8): $2 / $12. Similar price band, lower LMRank score.
- Qwen3.7 Max (9.0): $2.50 / $7.50, with cached input at $0.25. OpenRouter ran a 50% launch promotion.
- DeepSeek V4 Pro (9.0): $0.435 / $0.87 with DeepSeek's current 75% discount, plus open weights.
Qwen3.7 Max beats the US flagships on price, but DeepSeek V4 Pro beats it on price at the same score and ships weights. The case for Qwen over DeepSeek has to rest on your own evals and on Qwen's agentic tooling, not on price.
Trade-offs
- Output volume: at four times the median token count, the effective output price looks more like $30 per million than $7.50. That erases the gap with GPT-5.5 on reasoning-heavy tasks. Measure cost per completed task.
- Closed weights: no self-hosting, no fine-tuning outside Alibaba Cloud.
- Data residency: Alibaba Cloud Model Studio is the primary endpoint. Check whether that works for your compliance requirements before you look at benchmarks.
- Benchmark provenance: every number above is vendor-reported, and the comparison set is a generation behind.
Who should test it
- Teams running long-context document or codebase work where input dominates the bill. Here the $2.50 input price and $0.25 cached input matter more than verbosity.
- Teams whose agents are capped by step count rather than tokens, where a chattier model costs less than it appears.
Skip it if your workload is output-heavy reasoning; the verbosity makes it no cheaper than the models it is supposed to undercut.
Takeaway
Qwen3.7 Max is a credible near-frontier model at a low list price, not a hidden bargain. Run it on 50 real tasks, compare total spend against GPT-5.5 and DeepSeek V4 Pro, and let the per-task number decide.
Related: Cheapest Models · Longest-Context Models