Claude Sonnet 5.5 vs Qwen3.8 Max

vs

什么时候选哪个

纸面上这两款几乎处在同一位置:都接收文本和图像、返回文本,输入都是每百万 token $2,上下文都接近一百万 token(claude-sonnet-5-5 为 1000000,qwen3.8-max 为 983616)。差别在输出:qwen3.8-max 每百万收 $6,对方为 $10,便宜约 1.67x,单次回复空间也略大,为 131072 token。想要 Anthropic 更新的 2026-09-28 版本、以及 $0.1 对 $0.25 的缓存读取价时选 claude-sonnet-5-5;输出量大的工作选 qwen3.8-max。

Benchmark 成绩

Claude Sonnet 5.5:公布了 10 项成绩,但每一项测过的其他模型都太少,没法比较。

高于同类均值无更高分Qwen3.8 Max31 / 408 / 40
Claude Sonnet 5.5 Qwen3.8 Max 其他参测模型 同类均值 ★ 没有模型得分更高
SWE-Bench Pro
N/A
67.7%
AndroidBench
N/A
75.1%
Cybergym
N/A
78.5%
HealthBench
N/A
没有模型得分更高 60.2%
JobBench
N/A
53.4%
PLawBench
N/A
没有模型得分更高 73.2%
GPQA Diamond
N/A
92.6%
Agents' Last Exam
N/A
27%
Chartography No tools. Hydrogen and Terra: Google evaluation. Sonnet: system card
61.6%
N/A

供应商公布: Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Anthropic Anthropic Anthropic Anthropic ByteDance DeepSeek DeepSeek Google Google Google Google Google MiniMax Moonshot Moonshot OpenAI OpenAI OpenAI OpenAI OpenAI OpenAI OpenAI Tencent Z.ai Z.ai Z.ai Z.ai

价格

Claude Sonnet 5.5 Qwen3.8 Max Δ
输入 / 1M token $2 $2 =
输出 / 1M token $10 $6 1.7×
缓存读取 / 1M token $0.1 $0.25 0.4×
缓存写入 1.25x (5m) / 2x (1h) 1.25x -

价格取自构建时的实时目录,最新价格见各模型页面。

两个模型所处的位置:全部 86 个按同一单位计费的聊天模型的每 1M token 输入价分布(对数刻度)

能力

Claude Sonnet 5.5 Qwen3.8 Max
工具调用 是 是
思考控制 始终开启 是,但供应商未公布调节参数
结构化输出 是 是
提示词缓存 显式(由你标记前缀) 隐式 + 显式
缓存有效期 5m default, 1h option explicit: 5m, reset on hit
最小缓存前缀 512 个 token 1,024 个 token

规格

Claude Sonnet 5.5 Qwen3.8 Max
输入模态 文本 图像 文本 图像
输出模态 文本 文本
发布日期 2026-09-28 2026-08-03
知识截止日期 2026-06 -
上下文窗口 1,000,000 983,616
最大输出 128,000 131,072
思考参数 thinking.type -
可选值
thinking.type
  • adaptive (default)
  • between_tools
-
默认值 adaptive, effort high -

规格照录自各供应商的文档;供应商没有公布的项目,对应的行直接省略,不做推断。 完整来源: Claude Sonnet 5.5 · Qwen3.8 Max

同一条提示词,两个模型,经网关实测

提示词 Explain to a CFO, in exactly three sentences, why our LLM bill doubled after we enabled reasoning. 检查 恰好 3 句

Claude Sonnet 5.5 通过 · 3 sentences

Enabling reasoning makes the model "think" before it answers, and that internal thinking is generated as extra text (reasoning tokens) that we are billed for at the same rate as output tokens, even though users never see it. Output tokens also cost several times more than input tokens, so a single request that used to produce a short answer can now generate many times more billable text behind the scenes. Because our request volume and prompts didn't change, the doubling comes almost entirely from this hidden thinking, and we can bring it down by limiting reasoning effort, turning it off for simple tasks, and reserving it for the workloads where the accuracy gain justifies the cost.

输出 205 tok 延迟 3.6 s

Qwen3.8 Max 通过 · 3 sentences

Enabling reasoning makes the model produce additional hidden steps before responding, and those tokens are billable. It also tends to lengthen each interaction because the model works through more possibilities before settling on an answer. Therefore, the bill doubled mainly due to higher compute and token usage per request, not necessarily because the number of requests doubled.

输出 378 tok (+305 思考) 延迟 8.6 s

指令遵循(恰好三句,数得出来)、受众适配(对 CFO 说话的口吻),以及下方 token 计数暴露出的隐藏思考计费差额。

提示词 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. 检查 修复通过测试

Claude Sonnet 5.5 通过 · 8/8 cases

**The bug:** `cur` counts the number of *adjacent equal pairs* rather than the number of items in the run. A run of length k contains only k−1 such pairs, so every result is one too small. The counter also starts at 0 and the loop starts at index 1, so a single-element list never updates `best` and returns 0 instead of 1. (An empty list should return 0.) In addition, `cur` resets to 0 on a mismatch, when it should reset to 1, because the item at `xs[i]` begins a new run of length 1. The claim that it "misses runs ending at the last index" is really the same off-by-one: the final run is counted, but as one less than its true length. Because `best` is updated inside the loop on every extension, the last run isn't skipped. […]

输出 444 tok 延迟 6.7 s

Qwen3.8 Max 通过 · 8/8 cases

```python def longest_run(xs): if not xs: return 0 best = cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 else: cur = 1 best = max(best, cur) return best ``` The bug is that the original code starts `best` and `cur` at `0`, so it counts adjacent equal *transitions* rather than the number of items in the run. A run of length `n` has only `n - 1` equal-neighbor transitions, so single-element inputs return `0`, and runs that reach the end are undercounted by one. Initializing the current run to `1` for the first element, resetting it to `1` on a break, and updating `best` from that count fixes the off-by-one.

输出 1616 tok (+1411 思考) 延迟 34.7 s

修复是否真的正确(能运行)、解释的信息密度,以及在一个范围明确的任务上的 token 效率。

提示词 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. 检查 合法 JSON,严格符合 schema

Claude Sonnet 5.5 通过 · 5/5 fields, guidance null

Here is the extracted JSON: ```json { "company": "Meridian Robotics", "quarter": "September quarter", "revenue_usd_m": 412, "growth_pct": 11, "guidance": null } ``` Notes on the choices: - **quarter**: The text says "September quarter" without specifying a fiscal year or Q-number, so I kept the wording as stated. - **revenue_usd_m**: The text gives "412" for the topline, with no currency or unit stated explicitly. I've taken it as 412 million USD per the field definition, but the text does not confirm that, and the figure is constant-currency. - **growth_pct**: "Eleven and change" means slightly above 11, so 11 is an approximation. The exact figure isn't given. - **guidance**: Guidance is withheld pending the Q4 close, so no guidance is stated, hence null.

输出 300 tok 延迟 3.5 s

Qwen3.8 Max 通过 · 5/5 fields, guidance null

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

输出 1199 tok (+1141 思考) 延迟 24.4 s

是否严守 schema(不臆造字段)、能否顶住幻觉压力(原文明说暂不给出 guidance),以及结构化输出路径的差异。

提示词 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. 检查 120 词,0 个禁用词

Claude Sonnet 5.5 通过 · 122 words, 0 banned, 1 question

**Introducing Cross-Provider Prompt Caching** Your API gateway now caches prompts across every connected LLM provider, whether you route traffic to OpenAI, Anthropic, Google, or a self-hosted model. Here's how it works: the gateway fingerprints each prompt prefix, stores it centrally, and applies the right provider-specific caching mechanism automatically. When you switch providers or fail over during an outage, your cached context follows you. Early testing shows input token costs dropping by up to 70% on repeated system prompts and long documents, with latency improvements of 40% or more on cache hits. Tired of paying to send the same instructions thousands of times a day? Enable it with one configuration flag. Cache rules, TTLs, and per-team isolation are available in the dashboard today.

输出 266 tok 延迟 3.8 s

Qwen3.8 Max 通过 · 120 words, 0 banned, 1 question

Today, our API gateway adds prompt caching across major model providers. It stores prompts and responses in one fast cache layer. Teams can lower token spend, reduce latency, and repeat reliable answers. The feature supports OpenAI, Anthropic, Google, and Mistral through one configuration. You can set retention rules, scope access, and invalidate entries quickly. How does your team maintain consistent results during provider outages? Approved cached responses keep applications stable while fallback routes recover. The dashboard shows hit rates, savings, latency, and provider usage. Engineers receive audit trails for every cached prompt, enabling safer testing. Product managers can compare cost trends before and after cache adoption. Start with a small route, then safely expand caching to production traffic right now.

输出 2744 tok (+2591 思考) 延迟 46.3 s

是否守住约束(字数预算、禁用词表、只能有一个问句)、文风特征,以及长度控制。

改一行代码就能在两个模型之间切换

下方每个标签页里都有两个模型 ID,高亮的那两行是唯一要改的地方。端点不变,API key 不变,请求结构也不变。

from openai import OpenAI

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

resp = client.chat.completions.create(
    model="claude-sonnet-5-5",
    # model="qwen3.8-max",  # 取消注释此行,注释上一行
    messages=[{"role": "user", "content": "Summarize this diff"}],
    reasoning_effort="medium",
)
print(resp.choices[0].message.content)

获取 API key →

常见问题

Claude Sonnet 5.5 和 Qwen3.8 Max 哪个更便宜?

两者的「输入 / 1M token」同为 $2。其他费率与实际用量仍可能影响总费用。

不用分别集成两次,就能对 Claude Sonnet 5.5 和 Qwen3.8 Max 做 A/B 测试吗?

可以。两个模型走同一个 OpenAI 兼容端点,用同一个 API key,切换时只要改一行里的模型名,所以可以给两个模型各分一部分流量,直接对比账单。

Claude Sonnet 5.5 和 Qwen3.8 Max 支持提示词缓存吗?

支持。两个模型的缓存读取价都低于各自的输入价,所以前缀能反复命中缓存的负载,实际成本会比按官网价估算的低。具体的缓存读取价见上方价格表。

相关对比

我们的实测研究