Qwen 3.7 Max vs Plus: an honest split.

Quick answer: Plus for most seats, Max for a few.

For most workloads the answer is qwen3.7-plus. It launched on June 1, 2026 as the family's volume tier, at roughly one-sixth of Max's per-token input rate by launch coverage's math, and independent launch-season comparisons found the quality gap between the tiers narrow on knowledge tasks and widest on agent-specific benchmarks. Paying the flagship rate for traffic that Plus handles identically is the most common misrouting in this family. The seats Max genuinely earns: multi-step agentic plans, the hardest code generation, and long-horizon autonomous runs, the exact ground Alibaba positioned it for at the May 20, 2026 Cloud Summit announcement. If your workload has both kinds of requests, route by request rather than picking one id for everything; both tiers sit behind one key here, so the split costs nothing to implement.

The spec sheet, side by side.

Both tiers share the 3.7 text backbone, a 1M-token context window, and OpenAI-compatible serving. The differences that matter sit in modality, output ceiling, and rate. Figures below come from Alibaba's launch announcements and launch-season coverage and comparisons, as of July 2026; treat third-party benchmark deltas as launch-week snapshots rather than permanent truths.

Where the quality gap is real, per the launch-season evidence.

Independent comparisons published around the Plus GA measured the two tiers on shared benchmarks, and the pattern was consistent across write-ups: the widest gaps showed on agentic evaluations (one comparison put the Terminal-Bench Hard difference near 8.5 points), a small Max edge of about 2 points on SWE-Bench Pro, and near-parity on knowledge benchmarks (an MMLU-Pro gap nearer 1 point). The same coverage reported Max around 10% faster on text-only latency. Read those numbers as directional: they are third-party launch measurements, not vendor guarantees, and leaderboards move. But the shape they draw matches the vendor's own tiering, the flagship pulls away exactly where tasks are long, multi-step, and agentic, and it is the shape to test against your own workload rather than assume. The equally important half: on everyday completion work, extraction, summarization, drafting, classification, the measured gaps were small enough that the rate difference dominates. A tier that scores within a point or two while costing a fraction per token wins every seat where that point is not the product.

The modality difference, and a serving caveat.

The practical routing consequence: the modality question decides platform, the capability question decides tier. If vision is load-bearing for your product, verify the serving path first; if text is the workload, the tier decision reduces to the benchmark-vs-rate tradeoff above.

• Upstream, Plus is the multimodal tier: Alibaba's launch materials describe image and video understanding, GUI-agent grounding, and visual document work on the 3.7 backbone. Max is text-only.
• Through this catalog both ids are served as text chat-completions models; if your workload needs image input on qwen3.7-plus here, verify it with one small request before building on it, and treat text as the contract if the test fails plainly.
• The output ceilings differ upstream, roughly 64K for Max against 32K for Plus per launch spec sheets, which matters for single very long generations: full documents, large file rewrites, long transcripts.
• Context is symmetric: both serve 1M tokens here, so neither tier forces chunking earlier than the other; the bill, not the window, is the long-context constraint.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Which tier, for which workload.

When the A/B runs, judge three numbers per side: quality on your rubric, completion_tokens per completed task, and latency on interactive seats. A tier that wins two of three wins the seat; a tier that only wins a benchmark you did not run yourself wins nothing.

• Multi-step agent plans, hardest code generation, long autonomous runs: qwen3.7-max. This is the ground the launch benchmarks separate on.
• Extraction, summarization, chat, classification, high-volume agent steps: qwen3.7-plus. Near-parity quality at a volume rate.
• Single responses longer than ~32K tokens: qwen3.7-max, by the upstream output ceilings.
• Planner-executor loops: Max plans, Plus executes. Two strings, one key, one usage log.
• Pinned legacy workloads: qwen3.6-plus exists here for behavioral stability, priced above qwen3.7-plus, so treat it as a bridge until evals re-run on 3.7.
• Undecided: run the A/B below and let completion_tokens and your rubric decide.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APISROUTER_API_KEY"],
    base_url="https://api.apisrouter.com/v1",
)

PROMPT = "Plan and write the migration for this schema change: ..."

for model in ("qwen3.7-max", "qwen3.7-plus"):
    r = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": PROMPT}],
        max_tokens=4000,
    )
    print(model, r.usage.completion_tokens, "output tokens")

The rate math that decides most seats.

Work one representative request shape through both tiers and the routing usually writes itself. Take a typical agent step of 3,000 input tokens and 800 output tokens: at catalog rates the Plus request costs a small fraction of the Max request, and that ratio compounds linearly with volume. A pipeline running fifty thousand such steps a month pays the difference between the tiers as real money for capability it may never invoke. The inversion is equally real: a workload of two hundred hard planning requests a month costs little on either tier, and if Max's plan quality saves one human intervention a week, the flagship rate is noise. Rate matters where volume lives; capability matters where failures are expensive. Route accordingly, and let the usage log audit the decision monthly, since both tiers itemize per request under one key.

FAQ