A.I. Coding Models in a Nutshell - Brent Marquez | Web and Application Development

I wanted to create a short list of A.I. LLM models I’ve been using and summarize my opinions on them, particularly in the context of A.I. assisted coding. This list is obviously subjective and only based on my experience, and will inevitably go out of date unless I can keep up with the furious pace new models are being released at, but at the time of writing, these are some of the models I’ve been using with brief notes based on what I’ve observed in practice.

Primary harnesses used: Codex, Claude Code, Cursor CLI, Devin CLI, Antigravity CLI, Pi, OpenCode

Practices:

Reset of the context window and sessions with a handoff at 150k to 200k tokens
Generally I stick to medium reasoning/effort level
Occasional use with skills such as Matt Pococks’ (wayfinder, grill-me-with-docs, etc.)
Planning a feature or app/hard debugging: Use Largest Frontier Models. Execution of the plan: Use lower powered variant models. (Ex: Claude Fable for planning, Claude Opus or OpenAI GPT-6 Sol for execution)
For cost effective use, get subscriptions vs. pay-as-you-go API key based token usage.

The below notes and usage patterns are primarily driven by two things: 1) Managing cost and usage limits, 2) Speed of execution. Larger, more powerful models will obviously not only cost more, but eat up your usage limits faster, and they are also slower.

Frontier Models

Largest models

Anthropic Fable 5.1: This is a very capable and intelligent model which I only use for planning something big, like a complex feature or an entirely new application I want to build. Or if there is a nasty bug or debugging problem I have not been able to make headway with using other models. It is the “big guns” (i.e. more expensive) and I only bring it in when necessary.
Anthropic Opus 5.5: This is the workhorse and one of my go-to models for coding. It is a big improvement over Opus 5 and I have been happy working with it. I leave it on medium effort for most tasks and mostly get good results.
OpenAI GPT-6 Sol: Another workhorse, again left on medium reasoning level/effort. For me it is comparable to Opus 5.5, though I might prefer Opus a bit over it. People cite how thorough the GPT models are and sometimes I find it will over-engineer things and write unnecessary or nonsensical tests, so it requires a bit more guiding and watching.
OpenAI GPT-6 Astra: Flagship model for OpenAI, and I would reach for this just like I might reach for Fable 5.1. I’m not sure if they are exactly comparable (I default to Fable 5.1), but it is another set of “big guns” you can bring in for extraordinary tasks or planning something complex.
Google Gemini 4 Argon: Google’s latest frontier model. I have been impressed so far and have had good results with some UI rework and patching up I needed to do recently. It also seems pretty quick. Google’s models in general just feel faster than the other models to me. Gemini is also known for it’s multi-modal capability so you can send it images etc., though I’ve done that with other models as well.
Meta Muse Wonder: Meta’s latest frontier model and what I’ve been using it for more is actually chat because it seems to generate responses that are easier to read, less verbose and in more plain language than the other models in my experience so far.
SpaceXAI Grok 4.6: Grok is similar to Gemini Flash a bit in terms of speed, perhaps not quite as snappy, but it feels faster than a lot of other models. I go to it when I want to get medium complexity sized tasks tackled quickly. Things like UI work, a large refactor (moving and renaming things), etc. Occasionally I will use it for more complex features, but generally I stick to Opus 5.5/GPT-6 for that.

Smaller models

Claude Sonnet 5+: Most suggest to switch to this for smaller tasks instead of using Opus. In practice, I rarely use it. I mostly just stay on Opus 5.5 medium.
Claude Haiku: I have rarely used this model and have not been impressed when trying it. Generally I ignore it and no longer reach for it for anything. It is the cheapest of the Claude series, but noticeably less capable.
Gemini Flash 3.8: The latest Gemini Flash version, highly capable, fast and an enjoyable model to use. Not bad for everyday coding use, though I tend to use Opus 5.5/GPT-6 Sol or Grok 5.6 over it normally. The common theme for me with these Gemini models is that they are just noticeably faster than other models I use.
Gemini Flash 3.5: Blazing fast, I only mention it because if I have a very small refactor or tiny UI update I need to make, I use this older version to do it in lightning speed, sometimes even just for the fun of it.
GPT-6 Terra: I rarely use this variant of GPT, though it’s use case is to switch to it for less complicated tasks instead of using GPT-6 Sol. In practice, I generally stay on GPT-6 Sol.
GPT-6 Luna: If I switch to a lower model on Codex, it might be this one. It has a good reputation among users as being fairly capable for a lower powered model, very cheap and some also suggest switching it to a high reasoning/effort for a great bang for your buck. Generally, I find myself more or less staying on GPT-6 Sol these days for most tasks.

Open Models

I have not had as good an experience using Open Models as with the Frontier Models mentioned above, but some of them are decently capable, and in a lot of cases much cheaper to use comparatively.

Z.ai GLM 5.3 (has a flash variant as well): I am a fan of the GLM models. My first experience with GLM 5 was very positive and though I don’t use Open Models regularly like I do Frontier Models, if I’m going to use one, a GLM 5+ model is one of the first ones I’ll reach for. One time it did mangle a test file, which was disappointing, but on the whole, for coding in general GLM 5+ models are solid in my experience.
DeepSeek V4 Pro/V4.1 Flash: DeepSeek’s models are considered by many to be the best deal out there in terms of capability for low cost. I have had generally good experiences using their models, particularly the V4 variants in Pro and Flash mode. They are much cheaper than many other models and a good option to reach for when on a budget or you want to try to tackle a reasonably complex task for as cheap as possible.
MiniMax M3: This open model is considered one of the more capable out there, though in practice I have not had stellar results. In one instance, the model mangled a simple README file update, and I was not particularly impressed with the implementation code it produced. I do not use it very much anymore, but keep it in mind as a cheaper capable open model to reach for if needed.
MoonshotAI Kimi K3: Some say this is comparable to Fable 5, which is very impressive for an open model. I have had pretty good results using it to implement a very complex UI after planning and mocking it out. In practice, I will just use Fable, because this open model is atypically expensive and you’re not getting as great a “bang for your buck” vs. just using the proprietary model if you prefer it.
Qwen 3.8-27B: The 27 Billion parameter variant of one of Qwen’s latest open models. This is the first local model that can be hosted on consumer hardware that has impressed me. It is a huge leap forward for models which are small enough to be hosted privately by an average user not requiring resources only larger companies would have. I have been hosting it on a small AWS GPU box and while it cannot compare with frontier model capability, it is fine for smaller tasks and refactors or prototyping at zero cost and 100% privacy and data retention.

Further Resources

Be sure to check out my post on Adventures with A.I. Coding Agents for additional information and an overview of coding harnesses, best practices for security when using agents and usage examples.