Hello Inbox
Maximize inbox placement for email marketing campaigns
Start new thread

Which AI models are you using? And why?

by
•

The large AI labs (Anthropic, OpenAI, Grok, Gemini, etc) are constantly releasing new versions and improvements.

People on social media often advocate for one specific model claiming that it performs tasks better than others.

For Hello Inbox we're using Anthropic because we tested responses with other models and found it to give the best responses for our use case.

But I'm curious to hear from others building or using AI why they are using a specific model? and why?

  • Which model did you pick?

  • Why did you pick that model?

  • Why do you think it's better than others?

  • What tasks are you performing?

134 views

Add a comment

Replies

Best

Did you notice a big difference in response quality between Anthropic and Gemini?

 Yes, especially for reasoning. We analyze deliverability test reports and generate recommendations for our clients, and Anthropic's Claude Opus 4.5 and Sonnet 4.6 crafted the best responses for our use case. However, I tested different models back in April so it's possible things have evolved since then.

 That’s interesting. Reasoning quality can make a big difference for this kind of analysis. I’d be curious to see how the newer models compare if you test them again now.

 Yes, I'm curious as well. Problem is, time. It's moving so fast it's hard to keep up with all the newest releases. So I'm trying to figure out how often and how much time I should be spending testing the newest models. I also want to test the top open source models against Claude Opus to see how they compare.

I've been using Opus 4.7 throughout. Sometimes deepseek r1 and sonnet 3.7 for smaller tasks or research but to do a task they almost always hallucinate and the only one I can basically depend on is Opus. I have also been using gemini image banana pro for images; tried many different ones for graphics and none outputs the quality I need.

 Yes Opus seems to perform the best. I really want to test the latest open source models to see how they compare against Opus. BTW, DeepSeek released V4 Pro recently. Have you experimented with any other open source models besides DeepSeek R1?

 For image models you may find this useful

 OpenRouter does not yet have it as a model so not yet but the v3 pro is not bad neither. I find r1 more reliable for long tasks over various days.
Thanks for the image link. I will check it out

 You're welcome. LMK if you end up finding something better for images.

We're on GPT for the generation path in a flashcard tool I build, and the model was honestly the least interesting decision in it.

What moved quality was forcing structured output. Every call goes through one service with a JSON schema attached, so the model returns an array of objects and never formats anything itself. Before that we asked for a delimited text blob in the prompt and burned weeks on quoting and escaping bugs that were really the model improvising. Once serialization moved out of the prompt and into code, a lot of what had looked like model quality turned out to be prompt design.

The one place the model does still show is counting. Ask for 40 items, get 34. We over-request and truncate in code rather than trusting the number, and that behavior was about the same everywhere we tested.

 thanks for sharing. Yes, good prompt design is key. So if the behavior was about the same everywhere you tested what made you ultimately settle with ChatGPT? Cost?

 Cost is part of it, but it wasn't the only part.

It was the schema contract. We send the JSON schema with strict mode on, so the API itself rejects a response that doesn't match the shape, instead of us hoping the model complied and writing a parse-then-retry path for when it didn't. That deleted a whole class of retry from our code.

Strict mode rejects `minItems`, so you literally cannot ask the schema for "at least 40 items." That's why our count check lives in code, and it's the kind of thing you only learn by shipping.

Cost shows up a level down. Card generation is narrow and repetitive, so it runs on a small model and only the harder paths get the expensive one.

 ha interesting. Looks like Anthropic has a similar feature called Strict Tool Use. Going to look into that further. Thanks for the tip.

Does it know when a simple 2 line reply is enough?

 Yes, but prompt design is key. Including strict constraints within the prompt returned the results I wanted.

Gemini 3.5-flash for the vision work, Google Cloud Vision alongside it, and Gemini embeddings for anything that needs similarity. GPT-4.1-mini sits behind all of it as a fallback only.

But the honest answer to your question is that the model was not the decision that mattered, and I agree with Siarhei here. Sorti has to look at a screenshot, work out what it actually is, a recipe, a product, a place, and then get back to the original link. Two things moved quality far more than swapping models did.

First, splitting it into passes instead of one big prompt. A vision pass on the image, then a separate pass on whatever link is in it. The single prompt trying to do both was where almost all of our hallucinations lived.

Second, taking category naming away from the model entirely. We used to ask it to name the folder, which is how we ended up with Recipes, recipes, Food Recipes and Cooking all existing at the same time. Now the model just describes the thing, and we embed the category names and merge anything above a cosine threshold. Deterministic code fixed what better prompting never did.

Flash over a bigger model is mostly cost and latency at our volume. For classification the accuracy gap was small enough that the accuracy per dollar was not close.

 Interesting. Thanks for sharing.

For coding: I have tried Anthropic, Codex, and Cursor. While I've found things I like/dislike about each, I've realized its not just the models but the processes I've built up around the models that makes it truly work like a coding system. My skills/hooks/rules with Anthropic are robust, so I stick with it.

For my software: I'll evaluate a few models for each use case with a sample data set. I'll look at a mix of things like quality, price, speed and then pick. This has so far landed me at a mix of Sonnet, Opus, and Haiku for text-based work and Gemini models for images. I'm early in development, though, so I'm curious if other approaches have worked better as you scale!

 Good insight. I'm also using Anthropic's Claude Code for development and a mix of Sonnet, Opus, and Haiku for text-based work within our SaaS application.

Claude Code is extremely powerful. But it really does depend on how you use it. I'm also curious how it will adapt as we scale, but I think as long as we keep it modular, we'll be fine.

When building Hello Inbox, I started by using an existing enterprise framework called Symfony and I wrote a very detailed claude[.]md file.

Then I built each feature one at a time to keep the context window as small as possible.

Then I manually tested the feature, ran a security scan, linted, then pushed to git.

Then moved on to the next feature.

Claude Code performed extremely well with this workflow and tight constraints.

I think using an existing framework that is modular and scalable with the workflow I described, should make it possible to scale the app over time.

odd answer but we dont pick one. we ship as an mcp so it runs inside whatever the user already has, claude, cursor or codex.

what surprised us is the client matters more than the model. same tools and same prompts, but how much it checks with u before doing something changes a lot between them.

 ok, interesting. Do you know which client most of your customers are using? And why?

I mostly use Grok, Gemini, and ChatGPT, depending on the task.

For coding, I prefer models that give me direct, implementation-focused output with minimal unnecessary explanation or comments inside the code.

In my experience:

  • ChatGPT is strong for debugging, architecture, and structured problem-solving.

  • Gemini works well when I need to handle larger context or analyze a lot of information together.

  • Grok has been useful for fast iteration, coding ideas, and exploring different approaches.

I use Claude less often for coding because, for my workflow, it sometimes adds more explanation or code comments than I actually need.

I don’t think there is one “best” model overall. The best model depends heavily on the task, and I usually switch between them rather than committing to only one.