Convika

Русской версии пока нет — показана английская.

We gave 9 AI models the same landing page prompt: here's what they built

Quick answer

We sent one landing page prompt to 9 current AI models through OpenRouter on November 16, 2026, once each. Claude Opus 5.5 (57 seconds, $0.15) and GPT-6 Astra (65 seconds, $0.27) were fastest to a finished page that fit a phone and used only our facts; DeepSeek V4 Pro was cheapest ($0.05) but invented shop policies, and GLM 5.3 and Qwen3.8 Max ran out of tokens first. No model's form sent the email anywhere.

Publish my AI-made page free

Free plan: no credit card, 1 published page, 50 form entries a month.

On this page

On the morning of November 16, 2026, we wrote one prompt for a one-file landing page with a sign-up form and sent it, unchanged, to the current flagship model of 8 AI labs through OpenRouter's API. Anthropic got 2 runs, Claude Fable 5.2 and Claude Opus 5.5, because it names a different model for each kind of work. Each model ran once with the same settings, and the 2 that ran out of room got one more run with more room. One run is one sample, so read the results as what each model did that morning, not as a ranking. We opened every page the same way in Playwright's Chromium, at 1280 px and on a 390 px phone, and recorded what it cost, how long it took, whether it fit the phone, and what its form did.

The model names, IDs, and prices below change every few weeks. Everything here was checked on November 16, 2026.

The results

One run per model, max_tokens 20,000, everything else at each model's default; checked November 16, 2026, 08:34 to 09:00 JST. Cost is what OpenRouter charged for the run; seconds run from sending the request to the last byte of the reply.

ModelCostSecondsFits a 390 px phoneForm with an email fieldConsole errorsFile sizeOnly the facts we gave
GLM 5.3$0.036159No page: the limit ran out while it was still reasoning
DeepSeek V4 Pro$0.046111YesYes027.3 KBNo
Gemini 4 Carbon$0.07173YesYes016.6 KBNo
Qwen3.8 Max$0.121558No page: the reply stopped inside the CSS
Kimi K3.1 Pro$0.14397YesYes026.9 KBYes
Claude Opus 5.5$0.15157YesYes016.2 KBYes
Grok 4.7$0.232539YesYes014.3 KBYes
GPT-6 Astra$0.27165YesYes016.7 KBYes
Claude Fable 5.2$0.531117YesYes019.7 KBYes
GLM 5.3, second run, 64,000-token limit$0.00519YesYes010.3 KBNo
Qwen3.8 Max, second run, 64,000-token limit$0.167639No: 396 px wideYes033.4 KBNo

The 11 runs cost $1.77 in all.

  • 7 of 9 models finished a page. All 7 fit a 390 px phone with no sideways scrolling, logged no console errors, and had a labelled, required email field.
  • Finished pages took 57 seconds to 9 minutes. Claude Opus 5.5 and GPT-6 Astra sent the first line of HTML within 10 seconds. Grok 4.7 reasoned for over 8 minutes before its first line.
  • 4 models invented facts. DeepSeek, Gemini, and the second runs of GLM and Qwen added cancel rules, baking days, or loaf weights the bakery never gave.
  • No form sends the email anywhere. Each page's script shows a message instead. Only GPT-6 Astra's says the address wasn't saved.

The prompt

We wrote the prompt as a small business owner might, for an invented bakery. Every model got this text as one message, with no system prompt:

Make a landing page for my small business, as one HTML file.

The business: Crumb & Kettle, a small bakery in Portland, Oregon. We're starting a weekly bread club: one sourdough loaf every Saturday, picked up at the shop, $9 a week. The first loaves go out on Saturday, December 5, 2026. I want people to join the waitlist with their email.

What the page needs:
- At the top: a headline, one line about the bread club, and the sign-up form.
- The sign-up form: an email field (required) and a button.
- How it works, in 3 steps.
- The price.
- A short FAQ with 3 questions.
- A footer with the address (123 Example St, Portland, OR) and opening hours (Wednesday to Sunday, 7 am to 2 pm).

Rules:
- One self-contained HTML file: the CSS in a <style> tag, any JavaScript in a <script> tag. No frameworks and no build step.
- Mobile-first: it must work on a 390 px wide phone with no sideways scrolling, and look good on a desktop too.
- No external images. Use CSS, emoji, or inline SVG for any pictures.
- Reply with only the HTML file, starting with <!DOCTYPE html>. No explanation before or after it.

The 9 models, and why these

For each lab we took the newest model that the lab, or OpenRouter's model list where the lab doesn't say, calls its flagship or strongest, read on November 16, 2026. We took the base model, not a "Pro" reasoning mode, a faster-serving variant, or a batch price.

ModelOpenRouter IDPrice per million tokens, input / outputWhat the lab says
GPT-6 Astra (OpenAI)openai/gpt-6-astra$10 / $50"our flagship model for complex reasoning and coding"
Claude Fable 5.2 (Anthropic)anthropic/claude-fable-5.2$8 / $40"for demanding reasoning and long-horizon agentic work"
Claude Opus 5.5 (Anthropic)anthropic/claude-opus-5.5$4 / $20"start with Claude Opus 5.5 for most workloads"
Gemini 4 Carbon (Google)google/gemini-4-carbon$2 / $10"Our strongest reasoning model"
Grok 4.7 (xAI, listed as SpaceXAI)x-ai/grok-4.7$2 / $6"Our flagship model for code and everything else"
DeepSeek V4 Pro (DeepSeek)deepseek/deepseek-v4-pro-0813$0.111 / $4.44Its API serves 2 models, deepseek-flash and deepseek-v4-pro (DeepSeek-V4-Pro-0813); DeepSeek doesn't say "flagship"
Qwen3.8 Max (Alibaba's Qwen team)qwen/qwen3.8-max-0902$2 / $6Alibaba Cloud's recommended models list qwen4-max-preview and qwen3.8-max; OpenRouter didn't carry the Qwen 4 preview
Kimi K3.1 Pro (Moonshot AI)moonshotai/kimi-k3.1-pro$2 / $10"our flagship model for long-horizon coding and end-to-end knowledge work"
GLM 5.3 (Z.ai)z-ai/glm-5.3$0.084 / $3.60"Z.ai's latest flagship model"

Prices are OpenRouter's list prices that morning; OpenAI's, Anthropic's, and xAI's own pages showed the same numbers. OpenRouter picks a provider for each request, so the seconds include that provider's speed: GLM 5.3's two runs went to different providers.

How we ran it and measured each page

Every request had the same body: the prompt as one user message, max_tokens 20,000, streaming on, and each model's default reasoning and temperature. We kept every reply, the scripts, and OpenRouter's record of each run, and checked each charge against OpenRouter's generation record and the API key's usage.

Each page was served from a local web server and opened in Playwright's Chromium twice: at 1280 × 800, and as a phone at 390 × 844 with touch on, where a page without a viewport tag lays out at 980 px. The checks were the same for every page:

  • Renders: the page loads with no uncaught error and shows its text.
  • Fits a 390 px phone: the window and the page are both 390 px wide, so nothing scrolls sideways.
  • Form: a <form> with an <input type="email"> that is required and labelled, and a submit button. We pressed the button with the field empty, then with test@example.com, and wrote down what happened.
  • Console errors, file size, and the prompt's facts on the page: $9, December 5, 123 Example St, the hours, and Saturday.

We read every finished page's text ourselves for anything the prompt didn't say.

The 20,000-token limit cost 2 models their page. GLM 5.3 spent all 20,000 tokens reasoning and wrote nothing. Qwen3.8 Max reasoned for 13,407 tokens and ran out partway through its <style> block, so its file has no body. Grok 4.7 went past the limit instead: it returned 38,408 tokens, 33,184 of them reasoning. We then gave GLM and Qwen one more run each with a 64,000-token limit. Those second runs are a second sample at a different setting, so the table keeps them apart.

What each model built

Each picture shows the first screen at 1280 px on the left and on a 390 px phone on the right.

GPT-6 Astra

The Crumb & Kettle page GPT-6 Astra made: a large green serif headline, Good bread. A better Saturday., an email field with a dark green Join the waitlist button, and a drawn loaf with a red one loaf every week stamp. On the phone the form comes before the drawing.
GPT-6 Astra: $0.27, 65 seconds, 16.7 KB.

Well: the only form that tells the visitor the truth: "Online signup isn't open yet, and your email hasn't been saved." It used only the facts we gave, and it is the one page in sage and dark green. Badly: at $0.27 it cost 1.8 times what Claude Opus 5.5 charged for a page of about the same size, with the same facts.

Claude Fable 5.2

The page Claude Fable 5.2 made: a cream page with a brown serif headline, Fresh sourdough, every Saturday., an email field labelled Email address and a rust Join the waitlist button, and a drawn loaf with steam to the right. On the phone the form comes first.
Claude Fable 5.2: $0.53, 117 seconds, 19.7 KB.

Well: a visible "Email address" label, its own "Please enter your email address." on an empty submit, and only the facts we gave. Badly: it cost the most of any run, 3.5 times what Opus 5.5 charged for a page much like it.

Claude Opus 5.5

The page Claude Opus 5.5 made: a cream page with a dark red serif headline, Fresh sourdough, every Saturday., a white card holding the email field and a brown Join the waitlist button, and a drawn loaf with steam. On the phone the card comes before the drawing.
Claude Opus 5.5: $0.15, 57 seconds, 16.2 KB.

Well: the fastest finished page, with the same structure and the same discipline about facts as Fable 5.2. Badly: the plainest layout of the day, close to a template, down to the same steaming loaf Fable drew.

Gemini 4 Carbon

The page Gemini 4 Carbon made: a centered layout with a small loaf drawing, First pickup: Saturday, Dec 5, 2026 in small red capitals, a serif headline, Warm Sourdough, Waiting for You Every Saturday., and an email field beside a rust Join the Waitlist button.
Gemini 4 Carbon: $0.07, 73 seconds, 16.6 KB.

Well: the cheapest finished page from a US lab, in a clean centered layout. Badly: it invented Pacific Northwest organic flour, a 36-hour ferment, 48 hours' notice to skip a week, a 900 g loaf, and morning pastries. Its thank-you says "We saved test@example.com", and nothing was saved.

Grok 4.7

The page Grok 4.7 made: a large serif headline, One loaf. Every Saturday., with Saturday in red italics, a short line about the club to the right, and a white card holding Join the waitlist, an email field, and a red button.
Grok 4.7: $0.23, 539 seconds, 14.3 KB.

Well: the smallest file and the shortest page of the 7 finished at the shared limit. It used only our facts, wrote "Pickup at the shop — not delivery", and left the FAQ answers open instead of hiding them in accordions. Badly: 9 minutes and 33,184 reasoning tokens for that, past the 20,000-token limit. On the desktop the form card starts right under the headline, with no space.

DeepSeek V4 Pro

The page DeepSeek V4 Pro made: a centered layout with a Portland, Oregon tag, a bread emoji, a serif headline, Weekly Sourdough Bread Club, and an email field beside an amber Join the Waitlist button. On the phone the button spans the screen.
DeepSeek V4 Pro: $0.05, 111 seconds, 27.3 KB.

Well: the cheapest finished page at the shared limit, with every section the prompt asked for. Badly: it says pick-up is "any time after 8 am" when the shop opens at 7, calls the loaf "delivered to your hands", and adds cancel-anytime and food-pantry policies. It also wrapped its reply in a Markdown code fence, against the prompt, and made the largest file of the 7.

Kimi K3.1 Pro

The page Kimi K3.1 Pro made: a two-column hero with the headline Good bread. Every Saturday. in serif and red italics, small chips for the price, the loaf and the first pickup date, a white form card, and on the right a drawn loaf with a round price badge.
Kimi K3.1 Pro: $0.14, 97 seconds, 26.9 KB.

Well: the most illustrated page, a two-column hero with a drawn loaf and a $9 badge, using only the facts we gave. Badly: the longest page on the phone at the shared limit, 3,895 px from top to bottom.

Qwen3.8 Max

The page Qwen3.8 Max made on its second run: a yellow scrolling ticker under the header, a large serif headline, One loaf. Every Saturday., an email field with a yellow Join the waitlist button, a countdown to the first bake, and a drawn loaf beside a green kettle with a rotating round badge.
Qwen3.8 Max, second run with a 64,000-token limit: $0.17, 639 seconds, 33.4 KB. The first run, $0.12 and 558 seconds, left no page.

Well: the most elaborate page of the day, with a scrolling ticker, a live countdown to December 5, and its own error message for a missing address. Badly: the first run stopped inside the CSS. On the second, a rotating badge ends past the right edge, so the page is 396 px wide on a 390 px phone and slides sideways. It also invented a Friday bake, a 900 g loaf, and "gone by noon" for a shop open until 2 pm.

GLM 5.3

The page GLM 5.3 made on its second run: a centered serif headline, The Saturday Bread Club, a short paragraph with Saturday and the date in amber, an email field in a framed box with a brown Join the waitlist button, and a How it works list below.
GLM 5.3, second run with a 64,000-token limit: half a cent, 19 seconds, 10.3 KB. The first run, $0.04 and 159 seconds, left no page.

Well: the second run was the cheapest, fastest, and smallest page of the day. Badly: the first run used all 20,000 tokens reasoning and wrote nothing. The second invented a Friday bake, a 24-hour ferment, and "loaf numbers".

Claude vs GPT-6 vs Gemini vs Grok on this prompt

On this one prompt, Claude Opus 5.5 and GPT-6 Astra were the fastest, and both Claude models, GPT-6 Astra, and Grok 4.7 kept to the facts we gave. Grok took 9 minutes. Gemini 4 Carbon was quick and cheap, and filled the page with details the bakery never gave.

Claude Opus 5.5Claude Fable 5.2GPT-6 AstraGemini 4 CarbonGrok 4.7
Cost of this run$0.15$0.53$0.27$0.07$0.23
Seconds571176573539
Reasoning tokens6784,4831951,05933,184
Only the facts we gaveYesYesYesNoYes
What the form said on submitThanks, you're on the waitlistYou're on the waitlistYour email hasn't been savedWe saved your emailYou're on the list

GPT-6 vs Claude Opus 5.5: Opus 5.5 was 8 seconds faster and cost $0.12 less. Both pages kept to our facts and fit the phone. Astra's form was honest that it saves nothing; Opus 5.5's checked for an empty field with its own message.

DeepSeek, Qwen, Kimi and GLM on this prompt

Kimi K3.1 Pro was the one model from these 4 labs that finished at the shared limit using only our facts. DeepSeek V4 Pro made the cheapest finished page at the shared limit, and the one that most changed the facts we gave. Qwen3.8 Max and GLM 5.3 needed a higher limit to finish at all. On their second runs both made complete pages, and both added a Friday bake and a ferment time of their own.

Which model to use for what

If you wantIn this run
A page that says only what you told itClaude Opus 5.5, Claude Fable 5.2, GPT-6 Astra, Grok 4.7, Kimi K3.1 Pro
A finished page in about a minuteClaude Opus 5.5 (57 s), GPT-6 Astra (65 s), Gemini 4 Carbon (73 s)
The lowest cost for a finished pageDeepSeek V4 Pro ($0.05) or Gemini 4 Carbon ($0.07), then check every sentence
The smallest fileGrok 4.7 (14.3 KB)
The most visual pageKimi K3.1 Pro; Qwen3.8 Max with a higher limit, after you fix its width
A form that doesn't claim to save an emailGPT-6 Astra

Through the API, give a reasoning model room. With a 20,000-token limit, GLM 5.3 and Qwen3.8 Max returned nothing usable. If a reply stops with the finish reason length, raise the limit or ask for less CSS.

These results beside Arena's WebDev rankings, the best free model, and every way to put the page online: the best AI model for an HTML website.

Whichever model you use, check the page before you share it: what to check on an AI-made page before you share the link covers made-up facts and dead forms.

Put one of these pages online with a form that keeps entries

Every form in this test shows a message and keeps nothing: none has an action address that receives entries. To put a page live with a form that does keep them, publish it on Convika:

1. Save the model's reply as one HTML file

Save the reply as index.html. If the model wrapped it in a Markdown fence (DeepSeek did), keep only what's between the fences. If you'd like Convika's form where the model's form is, ask the model first: "Replace the form with <div data-convika-form></div>."

2. Drop it into the upload box

Drop the file into the upload box at the top of this page, or paste the code, and click Get my preview link. The preview turns scripts and forms off, so it's for checking the layout. The link works for 3 minutes.

3. Keep the page

Click Keep this page — free, no card and sign up. The page moves into your account as you uploaded it.

4. Add a form that keeps entries

On the page's screen, click Add a form to my page, choose the fields, and set the button text. Convika puts its form where data-convika-form is, or above the footer. Remove the model's own form so visitors don't fill in one that goes nowhere. Entries arrive under Forms, with CSV export; Free shows 50 entries a month and keeps the rest for Pro.

5. Publish

Click Publish page, choose Public, and click Next. The first time, click Publish free. The page goes live with the Convika bar at its foot; Pro ($12 a month) removes the bar.

For other ways to make a form keep entries, see how to make the sign-up form on your AI-made page actually work.

Frequently asked questions

Which AI model is best for making a landing page?

No single model was best in our test on November 16, 2026, and one run per model is one sample. Claude Opus 5.5 and GPT-6 Astra made a finished page that fit a phone and used only our facts in about a minute, for $0.15 and $0.27. DeepSeek V4 Pro made the cheapest finished page, at $0.05, but invented policies; Kimi K3.1 Pro drew the most visual page, and Grok 4.7 was faithful to the facts but took 9 minutes.

Is Claude or ChatGPT better at building websites?

On our one prompt, Claude Opus 5.5 finished 8 seconds sooner than GPT-6 Astra and cost $0.12 less, and both kept to the facts we gave. Astra's form was the only one that told visitors their email wasn't saved. We tested the models through the API, not the ChatGPT or Claude apps, whose settings can differ.

How much does it cost to make a landing page with an AI API?

Between half a cent and 53 cents a page in our runs, for pages of 10 to 33 KB. The price per token sets the range, and the model's reasoning tokens decide where a run lands in it: Grok 4.7 reasoned for 33,184 tokens and cost $0.23, while GPT-6 Astra reasoned for 195 and cost $0.27 at a much higher price per token.

Can DeepSeek, Qwen, Kimi, or GLM build a website?

Yes. Kimi K3.1 Pro and DeepSeek V4 Pro each made a complete page at a 20,000-token limit, and Qwen3.8 Max and GLM 5.3 did on a second run with a 64,000-token limit. DeepSeek, Qwen, and GLM added facts we never gave, so read their text before you publish.

Why did the AI stop before finishing my HTML?

It ran out of tokens. For most models in our test, reasoning counted against the output limit: Qwen3.8 Max ran out partway through its CSS, and GLM 5.3 used the whole limit thinking. Raise max_tokens, or in a chat, ask it to continue from where it stopped, or to make the page shorter.

Do the sign-up forms AI models write actually collect emails?

No. None of the forms in our test sent the address anywhere: each one ran a script that shows a message on the page. To collect entries, publish the page on Convika and click Add a form to my page: entries arrive under Forms. This guide explains why the form doesn't send and how to fix it.

Why does my AI-made page scroll sideways on a phone?

One element is wider than the screen. On Qwen3.8 Max's page it was a rotating round badge that ended about 6 px past a 390 px screen. Open the page in your browser's phone view, scroll right to find what sticks out, and ask the model to keep that element inside the screen width.

Sources

Read on November 16, 2026:

  • OpenRouter, model list (the API at openrouter.ai/api/v1/models, read at 08:30 JST): model IDs, list prices, and descriptions.
  • OpenAI, Models: "use GPT-6 Astra, our flagship model for complex reasoning and coding."
  • Anthropic, Models overview: "start with Claude Opus 5.5 for most workloads. Use Claude Fable 5.2 for demanding reasoning and long-horizon agentic work."
  • Google, Gemini API models: Gemini 4 Carbon, "Our strongest reasoning model."
  • xAI, Models: Grok 4.7, "Our flagship model for code and everything else."
  • DeepSeek, Models and pricing: deepseek-v4-pro is DeepSeek-V4-Pro-0813, thinking mode on by default.
  • Alibaba Cloud, Model Studio text generation models: qwen4-max-preview and qwen3.8-max.
  • Moonshot AI, Kimi API pricing: "Kimi K3.1 Pro is our flagship model for long-horizon coding and end-to-end knowledge work."
  • Z.ai, GLM-5.3: "GLM-5.3 is Z.ai's latest flagship model."

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

More ways to send your page

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

Share this article

Put the page your model made online

Upload the HTML file once. Convika hosts it, adds a form that keeps every entry, and counts visits.

Publish my AI-made page free

Free plan: no credit card, 1 published page, 50 form entries a month.