eStreetz
Claude, GPT, Gemini. What we actually reach for, and why.
Every few weeks a client asks which model is "the best". I've stopped answering that question directly, because the honest answer is "for what?", and that sounds evasive even when it isn't.
So here's how we actually choose at eStreetz right now. This is October 2026. Prices change monthly and new names show up even faster, so treat the numbers as a snapshot and check the pricing pages before you sign anything.
The short version
| Claude | OpenAI | ||
|---|---|---|---|
| Top model we use | Opus 5.5 | GPT-6 Astra | Gemini 4 Argon |
| Price (input / output, per 1M tokens) | $4 / $20 | $10 / $50 | $2 / $10 intro, then $4 / $20 |
| Mid tier | Sonnet 5.5, $2 / $10 | GPT-6.1 Sol, around $2 / $10 | Gemini 3.8 Flash |
| Cheap tier | Haiku 5.5, $0.10 / $0.50 | GPT-6 Luna, around $0.10 / $0.50 | Flash-Lite models |
| Context | 1M tokens | depends on tier | 1M tokens |
Gemini 4 Argon's launch price runs to the end of November. After that it doubles, according to Google's own docs. Plan for the second number.
What we use Claude for
Most of our agent work runs on Claude. EQO, our personal assistant product, runs on it, and so do most of the Uniwork agents.
The reason is boring: it follows instructions and it uses tools well. When an agent has eight tools and a long system prompt full of "don't send anything without approval", we need the model to actually respect that on turn 40, not just on turn 1. In our own testing Claude has been the most consistent at that.
It's also what we code with every day. Claude Code is open on most of our screens.
The thing to watch: newer Opus and Sonnet models have thinking on by default and you tune them with an effort setting. If you're migrating old code, re-test your effort level. The default on Opus 5.5 is medium, which caught one of our teammates off guard.
What we use GPT for
We reach for OpenAI when a client is already deep in that ecosystem (Azure contracts, existing assistants, fine-tunes), and for a couple of specific reasoning-heavy jobs where Sol has done well for us.
Astra is very capable. It's also the most expensive option in this table, and the long-context pricing goes up again past a certain prompt size. For most of what we build, it's hard to justify at that price.
What we use Gemini for
Long, messy, multimodal input. Video, scanned PDFs, a whole folder of screenshots. Gemini has been strong there for a while, and with Argon the price at launch is very hard to argue with.
We also like it for bulk work where we'd otherwise be choosing between a cheap model and a good one. Flash sits in a nice middle.
What actually decides it
Honestly, the model is rarely the deciding factor in whether a project works. What decides it:
- Do you have an eval? Even 30 real examples with expected answers. Without that, you're choosing models by vibes. We wrote about our pre-launch checklist separately.
- What does your client already pay for? Data agreements, cloud credits and procurement often matter more than benchmark points.
- Cost per finished task. A cheaper model that needs three retries isn't cheaper.
- Can you swap later? Keep the model behind one function in your code. We've switched providers mid-project twice this year and it took an afternoon each time, because of that one function.
My actual recommendation
If you're starting today and don't have a reason to pick otherwise: build on Claude Sonnet 5.5, keep Haiku for the small jobs, and run your eval against Gemini and GPT once a quarter to make sure you're not missing something.
Then stop reading model comparison posts, including this one, and ship the thing.
Written by Balaji B at eStreetz. Questions or disagreements: naveen@estreetz.com