CN Flagship Comparison · As of 2026-11-18

Qwen4-Max vs DeepSeek V4.1 Flash vs Kimi K3.1 Pro: Nov 2026 Coding & Cost Comparison

As of 2026-11-18, Alibaba Cloud's 2026-11-16 general-availability release of qwen4-max (2.4T total parameters, 71B active per token, 328B N-gram Embedding, 1M context with no long-context price doubling, official Alibaba Cloud Model Studio rate of ¥10 / $1.50 input and ¥30 / $4.50 output per 1M tokens, Work360 score 52.69 at $1.67 average task cost) joins the 2026-11-13 updated deepseek-v4.1-flash (DeepSeek-V4.1-Flash-1017, DeepSWE v1.1 at 77.0, Terminal-Bench 4.0 at 42.3), the 2026-11-05 kimi-k3.1-pro (1M context, 131,072 max output, vision support), and glm-5.3 as the core Chinese coding lineup. Below is a side-by-side breakdown of specs, official API rates, benchmarks, and how one QCode cr_ key calls all four.

Updated 2026-11-18

Switch models on the same endpoint and measure the cost per token yourself; check /models for callable models.

#qwen4-max (2.4T/71B)#deepseek-v4.1-flash#kimi-k3.1-pro (1M/131K)#glm-5.3 baseline

Four verified spec and cost highlights

$1.50 / $4.50

qwen4-max flat rate across 1M (¥10 / ¥30)

Official Alibaba Cloud Model Studio rate: flat pricing from 0 to 1,000,000 tokens with no long-context doubling — Beijing mainland node is ¥10 input, ¥30 output (including thinking chain), and ¥0.5 cache hit per 1M tokens; Singapore international node is $1.50 input, $4.50 output, and $0.075 cache hit.

52.69 / $1.67

qwen4-max Work360 score and task cost

Official benchmarks report qwen4-max scoring 52.69 on Work360 (up from 47.74 on qwen3.8-max) while cutting average single-task cost from $4.43 to $1.67 (a 62% reduction, with High-tier token consumption down ~35% and Terminal-Bench 4.0 token consumption down ~48% to 49%), ranking #6 on Code Arena and lifting Last Secure Code Bench pass rate from 22.0% to 31.8%.

77.0 / 42.3

DeepSeek-V4.1-Flash-1017 coding benchmarks

On 2026-11-13, DeepSeek updated online deepseek-flash to DeepSeek-V4.1-Flash-1017 at unchanged official rates ($0.15 input / $0.60 output off-peak, $0.30 input / $1.20 output peak per 1M tokens), posting 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0 with a 1M context window and 384K max output.

1M / 131,072

kimi-k3.1-pro specs and glm-5.3 reference

Moonshot's kimi-k3.1-pro launched on 2026-11-05 with a 1,048,576 (1M) context window, 131,072 max output, and native vision at Moonshot's official rate of $2.00 input, $10.00 output, and $0.20 cache hit per 1M tokens; Zhipu's glm-5.3 (1M context, 128K output) has an official rate of $1.40 input, $0.26 cache, and $4.40 output per 1M tokens.

Which engineering workloads each of the four models fits best

As of 2026-11-18, engineering teams choosing among Chinese flagship models for agentic coding evaluate four things: total per-task token cost, whether long contexts trigger tiered surcharges, multimodal input support, and protocol compatibility with local CLI tools. ① qwen4-max (released for pay-as-you-go on 2026-11-16) uses a sparse MoE architecture (2.4T total backbone parameters, 71B active per token, plus 328B N-gram Embedding parameters) supporting text, image, and video input across a 1,000,000-token window (non-thinking max input 991,808 and max output 131,072; thinking max input 983,616 and combined thinking-plus-answer max output 262,144); because its official Alibaba Cloud rate stays flat at ¥10 / ¥30 ($1.50 / $4.50 international) all the way to 1M tokens with tighter thinking chains, it fits multi-file refactoring and repository-scale reasoning. ② deepseek-v4.1-flash (552B MoE with 8B input / 16B output active parameters, 1M context, 384K max output) targets high throughput and low unit cost, reaching 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0 after the 2026-11-13 DeepSeek-V4.1-Flash-1017 update — ideal for high-concurrency sub-agents and test loops. ③ kimi-k3.1-pro (3T-class MoE, 1M context, 131,072 max output) pairs native vision with always-on thinking for UI screenshot debugging and deep code review. ④ glm-5.3 (1M context, 128K output, text-only) provides a steady coding baseline at Zhipu's official rate of $1.40 / $4.40. See /qwen4-max-api-pricing-guide for Qwen4-Max billing details and /ai-model-radar-2026 for live model availability.

What shipped across the three Chinese labs in mid-November 2026

Across the 12 days from November 5 to November 16, 2026, all three labs shipped production updates tailored to agent workflows: ① On 2026-11-05, Moonshot launched kimi-k3.1-pro with a 1M context window, 131,072 max output, native vision, and tool calling; ② On 2026-11-13, DeepSeek upgraded online deepseek-flash to DeepSeek-V4.1-Flash-1017, trimming task runtime and output length at the same official rates ($0.15 / $0.60 off-peak, $0.30 / $1.20 peak) while opening web search, file input, and URL reading across its Messages, Responses, and Chat Completions APIs; ③ On 2026-11-16, following its 2026-11-11 preview, Alibaba Cloud Model Studio launched the general-availability pay-as-you-go model ID qwen4-max with native support for all four protocols (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and DashScope) and cut implicit cache-hit pricing by 70% versus qwen3.8-max to ¥0.5 ($0.075) per 1M tokens.

November 2026 release timeline

2026-11-05

Moonshot launches kimi-k3.1-pro (1M context, 131,072 max output, native vision, official Moonshot rate $2.00 input, $10.00 output, $0.20 cache hit per 1M tokens); QCode adds kimi-k3.1-pro to /models on 2026-11-11.

2026-11-13

DeepSeek updates its online model to DeepSeek-V4.1-Flash-1017 at unchanged official rates ($0.15 / $0.60 off-peak, $0.30 / $1.20 peak), reporting 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0.

2026-11-16

Alibaba Cloud Model Studio launches general-availability pay-as-you-go qwen4-max (dated 2026-11-15 on the Qwen blog and 2026-11-16 in Model Studio release notes) with flat 1M pricing at ¥10 / ¥30 ($1.50 / $4.50), callable on QCode as of 2026-11-18.

Confirmed official data vs unannounced items

Confirmed (official vendor docs and pricing pages)

Verified as of 2026-11-18: ① Alibaba Cloud Model Studio docs: qwen4-max launched for pay-as-you-go on 2026-11-16 (the 2026-11-11 preview was Token Plan only and retires on 2026-11-25 at 24:00); sparse MoE with 2.4T backbone parameters, 71B active, and 328B N-gram Embedding; 1,000,000-token context (non-thinking max input 991,808 / max output 131,072; thinking max input 983,616 / max output 262,144); enable_thinking defaults to true, reasoning_effort accepts low, high (default), and xhigh; native support for 4 protocols (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, DashScope); official Alibaba Cloud API rate across 0 to 1M tokens is ¥10 input, ¥30 output, ¥12.5 5-minute cache write, ¥20 1-hour cache write, and ¥0.5 cache hit in mainland Beijing (Batch ¥5 input / ¥15 output), or $1.50 input, $4.50 output, $1.875 5-minute cache write, $3.00 1-hour cache write, and $0.075 cache hit on the Singapore international node ($1.413 / $4.24 on the international Beijing node); official benchmarks show Work360 at 52.69 with $1.67 average task cost (down 62% from 47.74 and $4.43 on qwen3.8-max), Code Arena #6, and Last Secure Code Bench pass rate at 31.8% (up from 22.0%). ② DeepSeek official docs: 2026-11-13 update to DeepSeek-V4.1-Flash-1017, 1M context, 384K max output, DeepSWE v1.1 at 77.0, Terminal-Bench 4.0 at 42.3, NL2Repo-Bench at 68.2, ProgramBench at 29.1, Agents' Last Exam at 35.4, HLE without tools at 42.6; official DeepSeek rate is $0.15 input, $0.60 output, $0.003 cache hit off-peak, and $0.30 input, $1.20 output, $0.006 cache hit during peak hours (UTC 01:00–04:00 and 06:00–10:00 Mon–Fri excluding China statutory holidays). ③ Moonshot and Zhipu official docs: kimi-k3.1-pro (2026-11-05, 1M context, 131,072 output, vision support) official rate is $2.00 input, $10.00 output, $0.20 cache hit; glm-5.3 (2026-08-18, 1M context, 128K output) official rate is $1.40 input, $0.26 cache, $4.40 output.

What is not announced or not supported

As of 2026-11-18: ① Alibaba Cloud's capability table marks model deployment and fine-tuning as unsupported for qwen4-max (closed-weight hosted flagship); while Alibaba Cloud stated smaller open-weight Qwen4 models will follow later, no release date, parameter size, or license has been announced, and there is no qwen4-plus, qwen4-flash, or qwen4-coder. ② DeepSeek has not announced a release date, model ID, or pricing for a V4.1 Pro model. ③ QCode serves only the production IDs listed on /models and does not offer kimi-k3.1-pro-highspeed or preview IDs.

Heavy reasoning & refactoring vs high-throughput execution

Repo-scale refactoring and multimodal 1M context: pick qwen4-max or kimi-k3.1-pro

When a task feeds hundreds of thousands of repository tokens at once or combines screenshots and video with code, qwen4-max keeps a flat rate across all 1,000,000 tokens without long-context doubling (official Alibaba Cloud rate fixed at ¥10 / ¥30, or $1.50 / $4.50), outputs up to 262,144 tokens in thinking mode, and averages $1.67 per task on Work360; when you want a second opinion from a 3T-class always-thinking vision model, switch in the same session to kimi-k3.1-pro (official Moonshot rate $2.00 / $10.00, 131,072 max output) or text-only glm-5.3 (official Zhipu rate $1.40 / $4.40).

High-concurrency sub-agents, terminal loops, and lowest unit cost: pick deepseek-v4.1-flash

When your workflow spawns parallel sub-agents, runs automated test loops, or needs up to 384K tokens of single-turn output, deepseek-v4.1-flash (powered by DeepSeek-V4.1-Flash-1017) scores 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0 at an official DeepSeek off-peak rate of $0.15 input / $0.60 output ($0.30 / $1.20 peak) and $0.003 / $0.006 cache hits — pairing naturally as the fast execution tier under qwen4-max.

Copy-paste setup to call all four models with one QCode cr_ key

① In Cline, OpenCode, WorkBuddy, or the OpenAI SDK: set Base URL to https://api.qcode.cc/openai/v1 (or https://asia.qcode.cc/openai/v1 in mainland China, with no trailing slash), put your QCode cr_ key in API Key, and set model to qwen4-max, deepseek-v4.1-flash, kimi-k3.1-pro, or glm-5.3. ② In Claude Code: set ANTHROPIC_BASE_URL to https://api.qcode.cc/api and ANTHROPIC_AUTH_TOKEN to the same cr_ key; inside any session, type /model qwen4-max, /model deepseek-v4.1-flash, /model kimi-k3.1-pro, or /model glm-5.3 to switch on the fly, or save the ID in the model field of ~/.claude/settings.json. ③ Thinking control: qwen4-max enables thinking by default (enable_thinking=true) and accepts reasoning_effort values of low, high (default), or xhigh; kimi-k3.1-pro runs in always-on thinking mode, so leave thinking enabled.

Switch across all four Chinese coding flagships on QCode

QCode supports qwen4-max, deepseek-v4.1-flash, deepseek-v4-pro, kimi-k3.1-pro, glm-5.3, glm-5.3-flash, qwen3.8-max, and qwen3.8-flash under the same cr_ API key you use for claude-fable-5-2 and gpt-6.2-sol. Billing is per token, with per-model rates on /models. Not ready for a monthly plan? Pay-as-you-go Packs start at $4 for a one-time quota you use within 7 days; if you already have a Personal Plan, add a booster pack when you need more quota.

Frequently asked questions

What are the official API rates for qwen4-max, deepseek-v4.1-flash, and kimi-k3.1-pro?

As of 2026-11-18, the official vendor rates per 1M tokens are: ① Alibaba Cloud Model Studio qwen4-max (flat rate from 0 to 1M): ¥10 input, ¥30 output (including thinking chain), and ¥0.5 cache hit in mainland Beijing; $1.50 input, $4.50 output, and $0.075 cache hit on the Singapore international node ($1.413 / $4.24 on the international Beijing node). ② DeepSeek official deepseek-flash (DeepSeek-V4.1-Flash-1017): off-peak $0.15 input, $0.60 output, $0.003 cache hit; peak (UTC 01:00–04:00 and 06:00–10:00 Mon–Fri excluding China statutory holidays) $0.30 input, $1.20 output, $0.006 cache hit. ③ Moonshot official kimi-k3.1-pro: $2.00 input, $10.00 output, $0.20 cache hit. ④ Zhipu official glm-5.3: $1.40 input, $0.26 cache, $4.40 output. QCode bills per token; check live per-model rates on /models.

Does qwen4-max double its price past 256K or 272K context tokens?

No. Under Alibaba Cloud Model Studio's 2026-11-16 general-availability pricing docs, qwen4-max uses a single flat rate across the entire 0 < Token <= 1,000,000 window (¥10 input / ¥30 output in mainland China, $1.50 input / $4.50 output on the Singapore international node) with no tiered surcharge; both explicit and implicit cache hits are ¥0.5 ($0.075) per 1M tokens, a 70% reduction on implicit cache versus qwen3.8-max.

Why did qwen4-max's average task cost on Work360 drop to $1.67?

According to Alibaba Cloud's published benchmark data, qwen4-max raised its Work360 score from 47.74 on qwen3.8-max to 52.69 while using about 35% fewer tokens at the High reasoning setting (and about 48% to 49% fewer tokens on Terminal-Bench 4.0); combined with lower cache pricing, average cost per task dropped from $4.43 to $1.67, a 62% reduction.

What coding benchmark gains came in the 2026-11-13 DeepSeek-V4.1-Flash-1017 update?

Per DeepSeek's official release notes, upgrading to DeepSeek-V4.1-Flash-1017 lifted DeepSWE v1.1 (Mini-SWE-V2 harness) from 74.2 to 77.0 and Terminal-Bench 4.0 (Claude Code harness) from 31.2 to 42.3, alongside 68.2 on NL2Repo-Bench, 29.1 on ProgramBench, and 35.4 on Agents' Last Exam, while reducing task runtime and output length at unchanged API rates.

Can I switch directly to qwen4-max or kimi-k3.1-pro inside Claude Code?

Yes. When Claude Code has ANTHROPIC_BASE_URL set to https://api.qcode.cc/api with your QCode cr_ key, you can run /model qwen4-max, /model deepseek-v4.1-flash, /model kimi-k3.1-pro, or /model glm-5.3 right inside the session without changing endpoints or restarting.

Are qwen4-max weights open-source, or is there a qwen4-coder / qwen4-plus?

As of 2026-11-18, qwen4-max is a closed-weight hosted flagship (Alibaba Cloud's official capability table lists model deployment and fine-tuning as unsupported); Alibaba Cloud announced that smaller open-weight Qwen4 models will be released later, without giving a date or license yet. Likewise, as of 2026-11-18 there is no official qwen4-plus, qwen4-flash, or qwen4-coder — use the production model ID qwen4-max.

Sources

Alibaba Cloud Model Studio official: qwen4-max model docs, pricing page, and 2026-11-16 release notes, plus the Qwen official blog (2026-11-15); DeepSeek official: API Change Log (2026-11-13 and 2026-10-17 DeepSeek-V4.1-Flash-1017 updates) and Models & Pricing page; Moonshot official: kimi-k3.1-pro release notes (2026-11-05) and pricing table; Zhipu official: GLM-5.3 model docs; QCode: qcode.cc/models and docs.qcode.cc. All checked on 2026-11-18.

Call Chinese and global coding flagships with one QCode key

Create a cr_ API key to run qwen4-max, deepseek-v4.1-flash, kimi-k3.1-pro, and glm-5.3 in Claude Code, Cline, and OpenCode — billed per token with per-model rates on /models.

Not sure where to start?

Try it for a week

Small Pack: $4 for $30 of quota, valid 7 days

Use it every day

Starter: $9.00/month with $5 of quota every day — just $0.060 per $1 of quota, vs about $0.113 for the Medium Pack