Comparing 11 different AI models

We just launched a partnership with OpenRouter that lets us offer two new pieces of functionality:

• First, your projects can use any model on OpenRouter through our AI Gateway. That means that if your own web app offers AI inference-based features to your end users, you now have a much wider selection of models to fit any task and budget.
• Second, we’re extending the selection of frontier coding models available for use via Agent Runners. Agent Runners is the chat prompt box you get within Netlify, which lets you build new projects from scratch or iterate on an existing one. The selection of models now includes much-hyped recent open models such as Kimi K3, GLM 5.2, and DeepSeek V4, available to everyone.

We call it Agent Runners because we run a full coding agent inside, not a pared-down one. Until now, we’ve supported Claude Agent, OpenAI Codex, and Gemini CLI which are optimized to run models from these providers.

We provide these agents with extra skills, and context about the current project, so that the agent will know exactly which Netlify capabilities are available for use (e.g., Netlify Database, the AI Gateway, or Identity), when to use them, and how. But to effectively drive a whole variety of new models, we’ve added the popular open-source OpenCode as a new choice of agent.

But with more choice come the inevitable questions: How do I know which model is right for me? Am I missing out on something that’s materially better, or more cost-effective (so I can do more with my credits), or is going to blow my mind like the internet says? There’s a lot of FOMO going around these days.

To provide you with some insights, here’s what we learned when running identical prompts across a range of models… all of which are now available for you to use today on Netlify.

You can see the results of all the models we tested on this site we created with the full report.

What we tested

Internally at Netlify, we use AXIS for automatically evaluating models, a tool that we’ve recently open-sourced.

We provide AXIS with a variety of test cases: prompts for building a new site and then iterating on it. We instruct AXIS on which agents and models to test these prompts, and define the checks that AXIS should then perform and score the generated site with.

These checks are very much focused on correct functionality of the generated site rather than its design, e.g.: does it use a database when a user’s needs call for it? Does it properly use Netlify Database in that case? In those cases where a simple static site will do, we also ensure that the generated site is not over-engineered, and no database is set up.

If a certain model is behind on its test scores, we don’t offer it in Agent Runners. If models too often fail at correctly applying one of our skills, or things do work but the credit cost seems inflated, then the problem is probably with the skill (in which case we optimize that skill).

But this time, we want to provide you with something much more immediately useful: when you go and build your dream using different models that each use wildly different amounts of credits, what do you get? What do the result look like?

We tested three relatively straightforward use-cases:

1. A site for a local coffee shop. LLMs just love making sites for local coffee shops! The initial prompt is simple, and a static site with no fancy database or the like will do. Then we do a follow-up prompt that asks for a simple option to reserve seats, and check how the model handled that.
2. A simple to-do list web app in which multiple users can view and add tasks. This calls for a simple design, but requires a shared database from the get-go. Then we ask to support an optional photo upload per item, and check if the model used the proper Netlify primitive.
3. A “What can I cook” web app that lets users enter what ingredients they have at home, and suggests a recipe using AI. The site itself is rather simple, but we want to check that the generated site correctly uses our AI Gateway to generate a recipe for the user.

For each of these cases, we’ll show you the look of the generated sites, comment on notable issues, and compare how many credits each took to generate. Of course, this is going to be a much more subjective test than our internal test suites, but it’s also going to be a very fun one. We’d love to know your opinion of the results!

All models were run with their default settings on Netlify. One notable mention is that we currently run GPT 5.6 Sol speicifically on low effort by default, giving you a more economical alternative to Opus that still provides pretty darn good results (as you’ll see below). However, the effort setting is now under your control, and our defaults may change with time.

This post is going to cover only the very first scenario: the static page for a coffee shop, while follow-up posts will focus on going beyond that simple use case. There is much to review even for this simple case, so let us begin.

Scenario #1: The local coffee shop

Here’s our first prompt:

Build a one-page site for a neighbourhood coffee shop: opening hours, the address, a short menu and a photo. Nothing on it changes unless I edit it myself.

The last sentence was added as a hint to the model that no fancy Content Management System is needed. Our default skills also include some UI design guidance, mainly to avoid known gotchas (e.g., the now-dreaded purple AI slop) and get the model to reason about the visual identity appropriate for the user’s ask. But beyond that, each model is free to go build what it thinks we’ll want.

Before we reveal what the sites looks like, here’s a table comparing the credit usage for each model we tested. Each model was run three times, and clicking any of the results will take you to the actual generated site!

That’s a pretty wide distribution, eh? Not only that: the Claude Opus average is heavily slanted upwards because one of its three runs spent a whopping 1,055 credits! (As a reminder, on the free plan you have 300 credits; on a Personal plan there’s 1,000 included credits; and with a Pro plan there’s 3,000 included credits. Additional credits packs for Pro are $10 for per 1,500 credits.)

The immediate question is then: is this Opus spend worth it? And what trade-offs do the other models offer? Let’s start digging in.

Claude Opus 5

Here’s the full page generated by that 1,055-credit run (about 4x more than any other run).

To be honest, I think it’s delightful, and full of detail in both its visual design (consider the “stamp like” element with the coffee bean in the center: that’s an actual text element that can be animated), and the custom map at the bottom. Dark mode works out of the box – go check out the live site in the links above.

Of course, we did not explicitly provide the model with any actual details about our coffee shop (well, except for it being a “neighbourhood” one, which is really steering all models in a certain direction). The design language is hip but perhaps cliche by now (take the two-font, two-color heading for example), but hey – we didn’t give it any other direction.

So, how did the other two runs by Opus go? (253 credits used on the left; 249 on the right)

Not bad either! Vector graphics actually require a lot of work from the models, and the examples above are pretty much on the frontier in terms of what LLMs currently are able to achieve (which is, to be honest, not in a very good place yet compared to image or text generation).

As to whether the first result is truly “4x better” or not, opinions might vary. But in all the tests I’ve done, Opus does have a tendency to run off with excessive credit usage (compared to its “typical” baseline) more than other models. It does not guarantee a worse or better outcome, though. It’s something that just happens pretty frequently.

Let’s look at some other models and then reflect on what we can learn.

Claude Sonnet 5

Here are our three contenders, at 143 credits on average (81 credits · 245 credits · 103 credits):

There’s still some delightful detail in each of these, just less so (and less content in general). The vector graphics is noticeably simpler and not really something you’d consider for a live site. This doesn’t say anything about this model’s ability to write complex code or answer philosophical questions, but we’re not asking for this here. At this price point, let’s see what OpenAI, Google and Kimi have to offer.

GPT 5.6 Sol (low effort)

What happens when we take OpenAI’s Opus-class model and ask it to spend a bit less time thinking?

(141 credits on average: 173 credits · 158 credits · 92 credits)

Looking into the results, I think OpenAI’s top-tier model in low effort mode wins over Anthropic’s mid-tier model when it comes to basic design intuition, at least in this scenario. There is more richness in content, and no funky vector shapes (though the images are a bit generic).

GPT 5.6 Terra

When we go one tier down in OpenAI’s offering (it’s Sol→Terra→Luna), will we see the same drop as the one we just witnessed when switching from Anthropic’s Opus to Sonnet?

Surprisingly, that’s not exactly the case: here it seems like Terra has a different visual language, and not a necessarily worse one. It does appear simpler content-wise. There are some visual glitches: a missing image in the left run, low-contrast text over an image in the middle one – but nothing super wrong.

(39 credits on average: 43 credits · 23 credits · 49 credits)

Up to this point, if I had a very vague idea of what design & language I’d like for a project, my personal inclination would be to run the same prompt with Opus 5 and GPT 5.6 Terra, and get two very different but worthwhile takes.

Gemini (3.6 Flash & 3.1 Pro)

These models are not of the same generation, and it shows: Gemini 3.6 Flash actually produced nicer results (or at least, more in line with other modern models) and used more credits compared to Gemini 3.1 Pro.

Here is what Gemini 3.1 Pro generated for 53 credits on average. I’m not even putting the links to the live site here, because there’s really nothing to see.

Yes, these are wholly separate runs. It did what we asked in the prompt, and really nothing more.

On the other hand, Gemini 3.6 Flash seems like a whole new generation, and used up 103 credits on average (109 credits · 91 credits · 111 credits). It also worked much harder on the content side of things. All models repeat themselves, but it seems like Gemini might repeat itself even more.

Kimi (K3 and K2.7 Code)

Ok, let us get to the open-weight models now. Starting with the latest Kimi K3, here is what we get (102 credits on average; 125 credits · 95 credits · 86 credits):

To be clear, Kimi K3 is marketed mostly as a frontier model for long-horizon agentic tasks, and various benchmarks and reviews confirm its prowess in that field. It was built to take on Fable 5 more than Opus 5. But in this narrow design-led task, it does not particularly shine among others. To really do this model justice, we’d need a wholly different set of prompts engineered for a complex web app, which we will cover in a follow-up post.

Going a big step back in model architecture to Kimi K2.7 Code, here is what we get for a very low credit average of just 19 credits:

Despite some hype about Kimi’s visual capabilities from around the K2.6 model launch, in terms of design or content there’s really not much to see here.

GLM 5.2

Let’s try this: look at these pages, ignore GLM’s love for maple, and try to estimate how many credits were used for each:

Here are the correct answers, from left to right: 15, 42, 24 (on average: 27). Surprisingly, these runs are – maple aside – very different, as if coming from a few different models. For the relatively low credit cost of GLM, it’s probably worthwhile to run it a few times before settling on what this model can do for you.

Note that being a text-only model that does not receive image inputs, GLM in its current 5.2 iteration cannot do something that Kimi models can: get screenshots from the user for inspiration, as in “this is the kind of design I’m looking for”.

DeepSeek V4 (V4 Pro and V4 Flash 0731)

V4 Pro is a bit older than the latest V4 Flash revision (also known as 0731). For about 47 credits, it does not provide inspiring results – especially compared to the mid-tier GPT 5.6 Terra model covered above, which sits at almost the same cost.

The middle run also has a broken image: the HTML file points to an image file that does not actually exist in the project, which is a lot less likely to occur nowadays with any of the commercial models from OpenAI, Anthropic, or Google.

V4 Flash 0731, on the other hand, is both newer and sets a new record here on how few credits it consumes.

For only 2.4 credits on average (3.4 credits · 1.3 credits · 2.5 credits), you get a mixture of results. Interestingly, the middle one doesn’t just look the most like what a mid-tier closed model might give you, but also feels the same in terms of language, and has actually consumed the least credits among all runs.

Interim conclusions, and what’s next

There are two important notes to make here:

First, for anything beyond a simple website or the initial ideation phase for a project, the question shifts from how nice the model design & copy is to:

• Does it know which platform features to use, when and how, to get the functionality you want? Can it store user data, use AI in your web app, and handle authentication and security?
• Does it rigorously validate its own work? Can it validate the frontend aspect of your project (that’s where image inputs become crucial)? Can it reliably find and fix issues based on feedback from you, and tell you when your own input is misleading or you’ve overlooked an important concern?

In the follow-up posts to this, we will start going into these questions, and (teaser) note some interesting differences in how models craft the project’s code.

My second note is that even considering just this design-and-copy-focused test that I covered, it’s important to consider how much ideation you want the model to come up with on its own. Currently, Opus will probably provide the most clever word games and sleekest design, but you don’t necessarily need it to. Of course, Opus will also perform relentless self-validation of its own work (it does not bill itself on good looks alone). But remember there’s certainly a higher-than-average credit cost attached to that.

Given a limited budget, would you prefer a turnkey solution that attempts to pre-plan and handle everything for you, or should you go with a simpler model and a more iterative approach, where you guide the model with follow-up prompts towards what you want? No option here is necessarily wrong.

I hope this post inspires you to test out different approaches, and judge for yourself the quality of results you get. We’re also pretty excited to share with you (very soon!) the results for more advanced web-app use-cases, where the Netlify platform capabilities really shine through.