DeepSeek vs Qwen vs Kimi for coding (November 2026): latest models, benchmarks, API prices, and where to run them
Quick answer
For coding agents, Kimi K3.1 Pro scores highest of the three on Terminal-Bench 4.0 in both Moonshot's and Qwen's own tables (65.7 and 67.2), at $2 / $10 per million tokens in / out. DeepSeek-V4.1-Flash-1017 costs $0.30 / $1.20 at peak and half that off-peak, and its 77.0 on DeepSWE v1.1 is the top score in DeepSeek's and Moonshot's tables. Qwen4-Max, on Qwen Cloud since November 16, 2026, costs $1.50 / $4.50 in Singapore and $1.413 / $4.24 in Beijing and Global, with no open weights, and Qwen3.8-27B (Apache 2.0, an 18 GB download in Ollama) is the one you run on your own computer.
Free plan: no credit card, 1 published page, 50 form entries a month.
On this page
- DeepSeek vs Qwen vs Kimi: the latest coding models side by side
- DeepSeek vs Qwen vs Kimi coding benchmarks: DeepSWE, Terminal-Bench, and SWE-bench Pro
- DeepSeek for coding: V4.1 Flash, the API, and coding agents
- Qwen for coding: Qwen4-Max, Qwen3.8, Qwen3-Coder-Next, and Qwen Code
- Kimi for coding: K3.1 Pro, the Kimi API, and Kimi Code
- Run it locally: Ollama, LM Studio, and vLLM
- Use it from Claude Code, Cursor, or Cline
- Which to pick
- Build and publish a page with it
- Frequently asked questions
- Sources
You want one of the three big Chinese model families behind your editor or coding agent, and you want to know which one codes best and what it costs. Each lab shipped a new model in the last 5 weeks: DeepSeek-V4.1-Flash-1017 on October 17, Kimi K3.1 Pro in October, and Qwen4-Max on November 16, 2026.
Every fact below comes from the lab's own model card, pricing page, docs, or launch post, read on November 16, 2026, between 20:16 and 20:30 Japan time. Benchmark numbers are each lab's own, with the lab named, because each runs the tests in its own setup. OpenRouter prices come from OpenRouter's own model list. For the wider field (GLM, MiniMax, licenses, and where each lab stores your data), see Chinese AI models in November 2026. For GPT-6, Claude, Gemini, and Grok, see the best AI models right now.
DeepSeek vs Qwen vs Kimi: the latest coding models side by side
DeepSeek vs Qwen vs Kimi for coding, checked November 16, 2026. Prices in US dollars per million tokens, input / output.
| DeepSeek | Qwen (Alibaba) | Kimi (Moonshot AI) | |
|---|---|---|---|
| Latest model | DeepSeek-V4.1-Flash-1017, API name deepseek-flash | Qwen4-Max, API name qwen4-max | Kimi K3.1 Pro, API name kimi-k3.1-pro |
| Released | October 17, 2026 | November 16, 2026 | October 2026; weights on Hugging Face November 2 |
| Context | 1M tokens, 384K max output | 1,000,000 tokens, 131,072 max output (262,144 with thinking) | 1,048,576 tokens |
| Open weights and license | Yes, MIT. 510 GB of weights | No: Qwen's Hugging Face page lists no Qwen4 weights on November 20, 2026. The newest open Qwen weights are Qwen3.8: the 2.4T-A95B model under the Qwen3.8-Max License, and Qwen3.8-27B under Apache 2.0 | Yes, under the Kimi K3.1 Pro License. 1.56 TB of weights |
| API price, on the lab's own page | $0.30 / $1.20 at peak, $0.15 / $0.60 off-peak | $1.50 / $4.50 in Singapore, $1.413 / $4.24 in Beijing and Global (checked November 20) | $2 / $10 |
| API price on OpenRouter | $0.15 / $0.60 on DeepSeek's own endpoint; 37 providers in all | $1.50 / $4.50, from Alibaba only | $2 / $10 on Moonshot's own endpoint; 18 providers in all |
| The lab's own coding scores | DeepSWE v1.1: 77.0. Terminal-Bench 4.0: 42.3 | DeepSWE v1.1: 75.7. Terminal-Bench 4.0: 48.5 | DeepSWE v1.1: 75.5. Terminal-Bench 4.0: 65.7 |
| The lab's own coding agent | DeepSeek Harness, in developer preview | Qwen Code: terminal, desktop app, VS Code, Zed, JetBrains | Kimi Code: desktop app, CLI, VS Code extension |
| Claude Code | Yes, through DeepSeek's Anthropic-format endpoint | Yes, through Qwen Cloud's Anthropic endpoint | Yes, through Kimi's Anthropic-compatible endpoint |
| Cursor | Listed as DeepSeek V4.1 Flash 1017, on Cursor Premium ($30 a month) and up | Not in Cursor's model list | Kimi K3 is listed; K3.1 Pro isn't |
| Cline (VS Code) | Built-in DeepSeek provider | Built-in Alibaba Qwen provider | Through OpenRouter or the OpenAI Compatible provider |
DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Its pricing page says "Off-peak rates are half of the peak rates", and weekends and Chinese public holidays are off-peak in full. Weight sizes are the totals of each Hugging Face repository's model files. Moonshot's posts give no launch day for K3.1 Pro: Kimi's API changelog lists it under October 2026, and OpenRouter added it on October 23.
DeepSeek vs Qwen vs Kimi coding benchmarks: DeepSWE, Terminal-Bench, and SWE-bench Pro
Two coding-agent tests appear in all three labs' current tables: DeepSWE v1.1 and Terminal-Bench 4.0. Each lab ran them its own way, and two of the labs also report a rival's number, so you can see where they agree.
Coding scores as each lab reports them, checked November 16, 2026. Higher is better. "Not listed" means that lab's table leaves the model out.
| Benchmark, reported by | DeepSeek-V4.1-Flash-1017 | Qwen4-Max | Kimi K3.1 Pro |
|---|---|---|---|
| DeepSWE v1.1, DeepSeek (October 17) | 77.0 | Not listed | Not listed |
| DeepSWE v1.1, Moonshot (K3.1 Pro model card) | 77.0 | Not listed | 75.5 |
| DeepSWE v1.1, Qwen (November 16) | Not listed | 75.7 | 75.5 |
| Terminal-Bench 4.0, DeepSeek | 42.3 | Not listed | Not listed |
| Terminal-Bench 4.0, Moonshot | 42.3 | Not listed | 65.7 |
| Terminal-Bench 4.0, Qwen | Not listed | 48.5 | 67.2 |
How each lab ran them:
- DeepSeek uses "the Claude Code harness for Terminal-Bench 4.0 and the Mini-SWE-V2 harness for DeepSWE v1.1."
- Moonshot ran K3.1 Pro in Claude Code for both tests, and took DeepSeek's numbers from DeepSeek's release.
- Qwen labels Moonshot's model "Kimi K3.1" in its chart. Qwen doesn't say which harness it used for these two tests.
What the numbers support: on DeepSWE v1.1, the three land within 1.5 points of each other (77.0, 75.7, 75.5). On Terminal-Bench 4.0, which grades agents on terminal tasks, Kimi K3.1 Pro leads the three in both tables that include it, at 65.7 and 67.2, against 48.5 for Qwen4-Max and 42.3 for DeepSeek.
Against US models, in the same tables. Moonshot's table gives Claude Opus 5.5 66.4 and GPT-6 Astra 58.2 on Terminal-Bench 4.0, beside Kimi K3.1 Pro's 65.7. Qwen's chart gives Claude Opus 5.5 59.6, GPT-6 Astra 59.1, and GPT-6.1 Sol 56.1. On DeepSWE v1.1, Qwen's chart gives Opus 5.5 74.2, GPT-6 Astra 74.1, and GPT-6.1 Sol 75.2. DeepSeek's own table compares with the older GPT-5.6 Sol (73.0 on DeepSWE v1.1) and Claude Opus 5 (74.0).
SWE-bench Pro. Of the three labs, only Qwen reports it for its current models: 67.7 for Qwen3.8-Max and 61.7 for Qwen3.8-27B, both run in the Claude Code harness. Qwen4-Max's chart doesn't include it, and DeepSeek's and Moonshot's cards don't list SWE-bench.
DeepSeek for coding: V4.1 Flash, the API, and coding agents
DeepSeek's API serves two models: deepseek-flash, which is DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro, which is DeepSeek-V4-Pro-0813. There is no separate coder model. DeepSeek's model card says the October 17 update delivers "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks" and "reduces task execution time and output length".
deepseek-flash (V4.1-Flash-1017) | deepseek-v4-pro (V4-Pro-0813) | |
|---|---|---|
| Input, cache miss | $0.30 peak, $0.15 off-peak | $1.32 peak, $0.66 off-peak |
| Input, cache hit | $0.006 peak, $0.003 off-peak | $0.044 peak, $0.022 off-peak |
| Output | $1.20 peak, $0.60 off-peak | $3.96 peak, $1.98 off-peak |
| Images in | Yes | "Not supported" |
| Fill-in-the-middle completion | Non-thinking mode only | Non-thinking mode only |
- Two API formats. The OpenAI format runs at
https://api.deepseek.com, and the Anthropic format athttps://api.deepseek.com/anthropic. DeepSeek's docs say: "If you use tools like Claude Code, GitHub Copilot, or OpenCode, you can use DeepSeek as the backend model directly — no code required." - Coding agents with a setup guide from DeepSeek: its own DeepSeek Harness, Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix, WorkBuddy/CodeBuddy, and Qoder. DeepSeek's GitHub list adds Cline, Kilo Code, Qwen Code, GitHub Copilot, and more.
- Tuned for agents. DeepSeek's changelog says of the October 17 model: "We specifically optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy, and other platforms."
- Paying. Fees come out of "your topped-up balance or granted balance".
Qwen for coding: Qwen4-Max, Qwen3.8, Qwen3-Coder-Next, and Qwen Code
Qwen gives you four ways to code: a new flagship on its API, open-weight general models, an open model built only for code, and its own coding agent.
Qwen4-Max came out on November 16, 2026. Qwen Cloud's page says it is "designed for coding, office productivity, frontend design, and complex agentic tasks." Qwen's launch post gives it "2.4T backbone parameters, 71B activated per token, and an additional 328B N-gram Embedding parameters." Against Qwen3.8-Max, Qwen reports about 48% fewer tokens on Terminal-Bench 4.0 at its highest effort setting. Its price is lower too: $1.50 / $4.50 in Singapore, against $2 / $6, and $1.413 / $4.24 in Beijing and Global, with no input-length tiers and a 1,000,000-token context. You call it through a Token Plan or the pay-as-you-go API on Qwen Cloud, in Singapore, Beijing, or the Global regions of Frankfurt, Hong Kong, Tokyo, and Virginia. Pay-as-you-go calls in Singapore come with 1 million free tokens. The weights aren't public; Qwen's post says "Smaller models in the family are coming to open source."
Qwen3.8 is the newest open-weight family. Qwen reports 67.7 on SWE-bench Pro for Qwen3.8-Max, in its Claude Code run; its open release is Qwen3.8-2.4T-A95B. Qwen3.8-27B is a 27B dense model under Apache 2.0, with 61.7 on SWE-bench Pro in Qwen's run, and Qwen Cloud lists it at $0.50 / $3.
Qwen3-Coder-Next, from January 30, 2026, is the newest Qwen model with "Coder" in its name: 80B parameters in total, 3B activated, a 256K context, under Apache 2.0. Qwen's card names "Claude Code, Qwen Code, Qoder, Kilo, Trae, Cline" among the tools it works with.
Qwen Code is Qwen's open-source coding agent. Install it with npm install -g @qwen-code/qwen-code@latest (Node.js 22 or later) and run qwen in your project. It runs in the terminal, as a desktop app, and in VS Code, Zed, and JetBrains. Its /auth menu offers 5 ways in:
- Qwen Account with Subscription: your Plus or Pro plan from chat.qwen.ai. Qwen Code's docs send you to chat.qwen.ai for the prices, and that page opens on a sign-in screen.
- QwenCloud: a Token Plan, or a pay-as-you-go API key.
- Alibaba ModelStudio: Coding Plan, Token Plan, or an API key.
- Third-party providers: DeepSeek, Kimi, Z.AI, MiniMax, OpenRouter, and others, with your own key.
- Custom Provider: a local server such as Ollama or vLLM.
Kimi for coding: K3.1 Pro, the Kimi API, and Kimi Code
Moonshot calls Kimi K3.1 Pro "our flagship model for long-horizon coding and end-to-end knowledge work". Its blog says: "K3.1 Pro covers the full range of coding work, from quick fixes in real codebases to full-project builds." The model card gives 2.8T parameters in total and 104B activated, a 1,048,576-token context, and text and image input.
Beyond the two shared tests, Moonshot reports 51.7 on SWE-Marathon v1.1 (DeepSeek-V4.1-Flash-1017: 40.0 in the same table) and 61.2 on its internal Kimi Code Bench 3.0 Long-horizon, up from Kimi K3's 55.95.
The Kimi API charges per token, with no subscription:
| Model | Input | Output | Cached input |
|---|---|---|---|
kimi-k3.1-pro | $2.00 | $10.00 | $0.20 |
kimi-k3.1-pro-highspeed ("same intelligence, about 2-3× the speed") | $4.00 | $20.00 | $0.20 |
kimi-k3 (the previous flagship) | $3.00 | $15.00 | $0.30 |
K3.1 Pro "is unlocked after a successful top-up (minimum $1)", and batch jobs cost 60% of the standard price. The API speaks both OpenAI and Anthropic formats, and Moonshot's blog says "Zero Data Retention is available for API customers."
Kimi Code is Moonshot's coding agent, as a desktop app, a CLI, and a VS Code extension. It runs on a Kimi membership, billed apart from the API: Kimi's docs say the two "are not interchangeable". Run /model in Kimi Code to pick K3.1 Pro.
Kimi's membership starts at Plus: $19 a month, or $15 a month billed yearly ($180). Three higher tiers add K3.1 Pro Highspeed and 2x or more agent credits; Kimi's membership page lists their prices. Moonshot's docs also give setup steps for Claude Code, Codex, OpenCode, OpenClaw, and Hermes Agent.
Run it locally: Ollama, LM Studio, and vLLM
On one computer, the Qwen models are the ones that fit. DeepSeek's and Kimi's open weights run on servers with 8 GPUs, and Ollama offers them as cloud models instead.
Running each model yourself, checked November 16, 2026.
| Model | Weights on Hugging Face | Ollama | What the lab or recipe names |
|---|---|---|---|
| Qwen3.8-27B (Apache 2.0) | 55.6 GB | ollama run qwen3.8: 18 GB, 256K context | Transformers, vLLM, SGLang, TokenSpeed. LM Studio lists it at 16 GB minimum system memory |
| Qwen3-Coder-Next (Apache 2.0) | 159.4 GB | ollama run qwen3-coder-next: 52 GB (q4_K_M), 85 GB (q8_0) | Ollama, LM Studio, MLX-LM, llama.cpp, and KTransformers for local use; SGLang or vLLM to serve it. LM Studio lists 42 GB minimum |
| DeepSeek-V4.1-Flash-1017 (MIT) | 510 GB | Cloud only: deepseek-v4.1-flash:cloud | DeepSeek's reference code splits the model across 8 GPUs |
| Kimi K3.1 Pro (Kimi K3.1 Pro License) | 1,561 GB | Cloud only: kimi-k3.1-pro:cloud | vLLM, SGLang, TokenSpeed. vLLM's recipe: 8 GPUs, as "two GB300 trays or one B300 node" |
Ollama's cloud tags send your requests to ollama.com, not your computer, and its pages list the labs' own prices: $0.30 / $1.20 at peak for DeepSeek, $2 / $10 for Kimi. Each Ollama model page also gives a one-line start for Claude Code, such as ollama launch claude --model qwen3.8.
To code with a local Qwen model in VS Code:
- Install Ollama and run
ollama run qwen3.8. The first run downloads 18 GB. - In VS Code, install the Cline extension. In Cline's settings, pick the Ollama provider, set the base URL to
http://localhost:11434, and chooseqwen3.8. - Turn on Use Compact Prompt under Cline Settings → Features, as Cline's local-model guide says.
Cline's guide matches memory to models: 16 to 32 GB for "Small/quantized models", 32 to 64 GB for "Mid-size coding models", and 64 GB or more for "Larger models and bigger context windows". Qwen Code reads a local Ollama or vLLM server through its Custom Provider. Cursor's own-key page lists OpenAI, Anthropic, Google, Azure OpenAI, and AWS Bedrock, so a model on your own computer goes through Cline or Qwen Code.
Use it from Claude Code, Cursor, or Cline
Claude Code. All three labs publish an Anthropic-format endpoint and a setup page for Claude Code. Set the variables, then run claude in your project.
DeepSeek's setup (its guide also sets the Opus, Sonnet, Haiku, and subagent models to deepseek-flash):
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API key>
export ANTHROPIC_MODEL=deepseek-flash[1m]
Qwen Cloud's pay-as-you-go setup, in Singapore (its guide also sets the Haiku, Sonnet, and subagent models). Qwen's docs say: "To use qwen4-max, set ANTHROPIC_DEFAULT_OPUS_MODEL to qwen4-max and keep CLAUDE_CODE_MAX_CONTEXT_TOKENS at 983616."
export ANTHROPIC_AUTH_TOKEN=<your Qwen Cloud API key>
export ANTHROPIC_BASE_URL=https://ap-southeast-1.maas.qwencloudapi.com/apps/anthropic
export ANTHROPIC_MODEL=qwen4-max
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen4-max
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=983616
Kimi's setup goes in the env block of ~/.claude/settings.json, with ANTHROPIC_BASE_URL set to https://api.moonshot.ai/anthropic and your Kimi key as ANTHROPIC_AUTH_TOKEN. Kimi's guide sets every model tier to kimi-k3[1m] and tests the key with a call to kimi-k3.1-pro.
When Claude Code asks for a Claude model by name, DeepSeek maps it: "Models starting with claude-opus are mapped to deepseek-v4-pro", and Sonnet and Haiku names go to deepseek-flash.
Cursor. Cursor's model list includes DeepSeek V4.1 Flash 1017 at $0.30 / $1.20, Kimi K3 at $3 / $15, and Kimi K2.7 Code, all hidden by default. Third-party models come with Cursor Premium ($30 a month) and up, at API list price from prepaid credit. No Qwen model is in the list, and neither is Kimi K3.1 Pro.
Cline. Cline has a DeepSeek provider and an Alibaba Qwen provider (choose the International or China region), and OpenRouter and OpenAI Compatible providers for anything else, Kimi included. DeepSeek's own Cline guide also shows the OpenAI Compatible route, with the base URL https://api.deepseek.com.
Codex and OpenCode. DeepSeek, Qwen Cloud, and Kimi each publish setup pages for both.
Which to pick
| If you want | Pick | Why, from the labs' own pages |
|---|---|---|
| The highest coding-agent score of the three | Kimi K3.1 Pro, $2 / $10 | 65.7 on Terminal-Bench 4.0 in Moonshot's table, 67.2 in Qwen's |
| The lowest price for long agent runs | DeepSeek V4.1 Flash, $0.30 / $1.20 at peak, $0.15 / $0.60 off-peak | 77.0 on DeepSWE v1.1, the top score in DeepSeek's and Moonshot's tables |
| The newest model, at a middle price | Qwen4-Max, $1.50 / $4.50 | Released November 16; 75.7 on DeepSWE v1.1 in Qwen's chart |
| A model on your own laptop | Qwen3.8-27B | Apache 2.0; an 18 GB download in Ollama; 61.7 on SWE-bench Pro in Qwen's run |
| A flagship's weights under a standard open-source license | DeepSeek-V4.1-Flash-1017 | MIT. Kimi K3.1 Pro's and Qwen3.8-Max's weights carry the labs' own licenses. DeepSeek's reference code runs it on 8 GPUs |
| A monthly plan instead of per-token bills | Kimi Code from $19 a month, or Qwen Code with a Qwen Plus or Pro plan | Kimi: "one subscription covers desktop, terminal, and IDE" |
| One key for all three | OpenRouter | DeepSeek V4.1 Flash, Qwen4-Max, and Kimi K3.1 Pro, each on its lab's own endpoint, plus other providers |
Before you move a team over, run your own 10 hardest tasks through 2 of them in the agent you already use. All three plug into Claude Code, so each test is a few environment variables away.
Build and publish a page with it
Ask any of the three for one HTML file, and the agent writes the page. In our test of 9 models with one landing page prompt, run the same day, Kimi K3.1 Pro made a complete page that fit a phone screen in 97 seconds, for $0.14. To put such a file online, drop it into Convika's publish tool for a preview link without an account; publish it from a free account to keep it at one address, and Add a form to my page adds a form that keeps the entries people send.
Frequently asked questions
Which is better for coding, DeepSeek, Qwen, or Kimi?
Kimi K3.1 Pro scores highest of the three on Terminal-Bench 4.0, a coding-agent test, in both Moonshot's table (65.7) and Qwen's (67.2). On DeepSWE v1.1, the three are within 1.5 points: DeepSeek V4.1 Flash 77.0, Qwen4-Max 75.7, and Kimi K3.1 Pro 75.5. DeepSeek costs the least, at $0.30 / $1.20 per million tokens at peak, against $1.50 / $4.50 for Qwen4-Max in Singapore and $2 / $10 for Kimi K3.1 Pro.
What is the latest DeepSeek model for coding?
DeepSeek-V4.1-Flash-1017, released October 17, 2026, which the API serves as deepseek-flash. DeepSeek reports 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0. It costs $0.30 / $1.20 per million tokens at peak and half that off-peak, with MIT-licensed weights. The larger deepseek-v4-pro costs $1.32 / $3.96 at peak.
What is the latest Qwen coding model? Is there a new Qwen3 Coder?
The latest is Qwen4-Max, on Qwen Cloud since November 16, 2026, through a Token Plan or the pay-as-you-go API. It costs $1.50 / $4.50 per million tokens in Singapore and $1.413 / $4.24 in Beijing and Global, with a 1,000,000-token context, and its weights aren't on Hugging Face. The newest model with "Coder" in its name is Qwen3-Coder-Next, from January 30, 2026: 80B parameters, 3B activated, under Apache 2.0. The newest open-weight general models are Qwen3.8-Max (as Qwen3.8-2.4T-A95B) and Qwen3.8-27B.
What is Kimi Code, and how much does it cost?
Kimi Code is Moonshot's coding agent, as a desktop app, a CLI, and a VS Code extension, running Kimi K3.1 Pro. It comes with a Kimi membership, from Plus at $19 a month, or $15 a month billed yearly. Three higher tiers add K3.1 Pro Highspeed and 2x or more agent credits. The Kimi API is billed separately, at $2 / $10 per million tokens for K3.1 Pro.
Can I run DeepSeek or Qwen locally with Ollama?
Qwen, yes: ollama run qwen3.8 downloads Qwen3.8-27B as an 18 GB file, and ollama run qwen3-coder-next downloads 52 GB. DeepSeek V4.1 Flash and Kimi K3.1 Pro are on Ollama as cloud models only, served from ollama.com. Their full weights are 510 GB and 1.56 TB, and DeepSeek's reference code and the vLLM recipe Moonshot links both use 8 GPUs.
How do I connect a local model to VS Code or Cursor?
In VS Code, use Cline: choose its Ollama provider, set the base URL to http://localhost:11434, and pick the model you pulled. Qwen Code also reads a local Ollama or vLLM server. Cursor's own-key page lists OpenAI, Anthropic, Google, Azure OpenAI, and AWS Bedrock, so use Cline or Qwen Code for a local model.
Can I use DeepSeek, Qwen, or Kimi in Claude Code?
Yes, all three. Point ANTHROPIC_BASE_URL at https://api.deepseek.com/anthropic for DeepSeek, https://ap-southeast-1.maas.qwencloudapi.com/apps/anthropic for Qwen Cloud in Singapore, or https://api.moonshot.ai/anthropic for Kimi, and put that lab's API key in ANTHROPIC_AUTH_TOKEN. Each lab's setup page names the model variables to set.
Can I use Kimi or DeepSeek in Cursor?
Yes, on Cursor Premium ($30 a month) and up, which adds third-party models at API list price. Cursor's list includes DeepSeek V4.1 Flash 1017 at $0.30 / $1.20 and Kimi K3 at $3 / $15. Kimi K3.1 Pro and the Qwen models aren't in Cursor's list.
Which is cheapest for coding: DeepSeek, Qwen, or Kimi?
DeepSeek V4.1 Flash: $0.30 / $1.20 per million tokens at peak, and $0.15 / $0.60 off-peak, which includes weekends and Chinese public holidays. Cache hits cost $0.006 at peak. For no API bill at all, run Qwen3.8-27B on your own computer under its Apache 2.0 license.
Sources
Each page was read on November 16, 2026, between 20:16 and 20:30 Japan time (11:16 to 11:30 UTC); dates in brackets are the page's own. Pages that draw their tables in the browser were read in a headless browser.
- DeepSeek: Models & Pricing, Your First API Call, Change Log (October 17 and November 13, 2026), Integrate with Claude Code, Integrate with OpenCode, DeepSeek-V4.1-Flash-1017 on Hugging Face and its inference guide, Awesome DeepSeek Agent and its Cline guide
- Qwen: Qwen4-Max: Your New Teammate Has Arrived (November 16, 2026), Qwen4-Max on Qwen Cloud, Latest model: Qwen4-Max, Qwen3.8-Max and Qwen3.8-27B on Qwen Cloud, Claude Code on Qwen Cloud, Qwen3.8-2.4T-A95B, Qwen3.8-27B, and Qwen3-Coder-Next on Hugging Face, Qwen Code and its authentication guide
- Kimi: Kimi K3.1 Pro post, Kimi K3.1 Pro on Hugging Face and its license, API pricing, model list, K3.1 Pro quickstart, Use Kimi in Claude Code, Compare with Other Kimi Products, platform changelog, Kimi Code, Kimi plans
- Tools and routers: OpenRouter models API and its endpoint lists for DeepSeek V4.1 Flash 1017, Qwen4 Max, and Kimi K3.1 Pro, Cursor models and pricing, Cursor: bring your own API key, Cline: DeepSeek, Cline: Alibaba Qwen, Cline: local models, Ollama's qwen3.8, qwen3-coder-next, deepseek-v4.1-flash, and kimi-k3.1-pro pages, LM Studio's Qwen3.8 27B and Qwen3 Coder Next pages, vLLM recipe for Kimi K3.1 Pro
Next step
How to password protect an HTML page or website
A JavaScript password script leaves the password in the page source. To password protect an HTML page or link, encrypt the file or let a server check it.
Related articles
- Chinese AI models in November 2026: the latest from DeepSeek, Qwen, Kimi, GLM, and MiniMaxEvery major Chinese AI model, from DeepSeek-V4.1 and Qwen4-Max to Kimi K3.1 Pro, GLM-5.3, and MiniMax M3.1: open weights, license, API price, and where to use them.
- The best AI models right now (November 2026): GPT-6, Claude, Gemini, and Grok comparedGPT-6 Astra, Claude Fable 5.2 and Opus 5.5, Gemini 4 Carbon, and Grok 4.7 compared: release dates, context, API prices, plans, and each lab's own benchmarks.
- We gave 9 AI models the same landing page prompt: here's what they builtOne landing page prompt, sent once to GPT-6 Astra, Claude, Gemini 4, Grok 4.7, DeepSeek, Qwen, Kimi and GLM on Nov 16, 2026. Cost, time, phone test.
- The best AI model for a one-page HTML website (November 2026), and how to put the page onlineClaude Opus 5.5 and Fable 5.2 top Arena's WebDev board; Opus built our test page in 57 seconds for $0.15. Put it online: drop the .html file on Convika.
- How to deploy and publish a website from Claude CodeYes, you can publish a website from Claude Code, but a claude.ai artifact can't keep signups. Deploy with Convika to add a form, your domain, and visit stats.
- DeepSeek HTML: run it in the chat, then publish it as a page anyone can openDeepSeek's Run button plays HTML beside the chat and Download saves an .html file, but DeepSeek won't host it. Paste the code here for a link anyone can open.
Skipbox
Cheap & fast GitHub Actions runners, unlimited for $39 a month
Your own machine keeps Docker layers and package caches between jobs. Switch by changing one runs-on line.
Start a free Skipbox trial