The Actual Ratio Today: 90% 5.6sol (Medium-High), 10% Antigravity — And Eagerly Awaiting GPT-6 Astra

Back in June, I shared how I split my development workflow between Antigravity for quick tasks and Codex (Sol model) for complex logic. At the time, that division felt like a relatively balanced two-track approach.

Over the past several months of intense, daily production work, practical experience continued to reshape the routine. Not out of dogmatic preference, but pure pragmatic efficiency: when a model makes fewer mistakes and delivers dependable code on the first attempt, you naturally end up delegating more and more of your workload to it.

Today, my actual workflow split stands at roughly 90% 5.6sol (medium-high reasoning) and 10% Antigravity.

Here is the real reasoning behind those numbers, and why I am watching the horizon for the release of GPT-6 Astra.

Why 5.6sol (Medium-High) Took Over 90% of the Work

When you are no longer building greenfield starter templates or simple landing pages, but maintaining and scaling mature, multi-thousand-line systems (such as VisaAlbum's document and image pipelines, or SabAI's stateful consular rules), the nature of coding changes:

"Simple" tasks aren't actually simple anymore: Adding a single new field to an existing flow doesn't just mean adding a JSX input; it touches relational database schemas, validation logic, server actions, and type contracts across multiple files. If a lighter model cuts corners here, you spend twice as much time tracking down subtle edge-case regressions later.
5.6sol at medium-high reasoning hits the sweet spot:
- Running at medium or high reasoning effort, the model doesn't rush. It methodically considers dependencies, boundary conditions, and AST-level invariants before outputting code.
- There is virtually no cleanup required: no dropped imports, no relaxed types, and no skipped edge-case branches.
Net velocity is higher: You might wait a few seconds longer for the model to think through the prompt compared to an instantaneous auto-complete, but because the code is correct on the first attempt, the total cycle time from intent to verified deployment is significantly shorter.

For these reasons, roughly 90% of my daily heavy lifting is now powered by 5.6sol with medium-high reasoning.

The Remaining 10%: What Antigravity Still Handles

Antigravity remains an active part of my setup, but its domain has narrowed to where its specific strength shines: instant speed on minimal-complexity tasks:

Quick UI and styling adjustments: Tweaking spacing, swapping a Tailwind color class, or fine-tuning button states.
Trivial copy and configuration updates: Adding a line to an i18n dictionary or verifying an environment variable.
Fast, routine inspections: Glancing at a file or checking terminal output when responsiveness is more important than deep contextual reasoning.

For these quick adjustments, firing up a heavy reasoning model is unnecessary; Antigravity responds instantly and keeps momentum moving without friction.

Looking Ahead: Awaiting GPT-6 Astra

The rate of progress in AI models shows no sign of slowing down. While I am thoroughly satisfied with my current 5.6sol (medium-high) setup, I hold no rigid brand loyalties.

I am particularly curious about the upcoming GPT-6 Astra. Several key questions will define its real-world utility:

Will it push multi-step autonomous reasoning even further while bringing response latency down?
How will it handle real-time, multimodal feedback loops across full application lifecycles?
Will it shift the current 90/10 split once again, or establish an entirely new capability tier in the daily toolkit?

As soon as Astra becomes available, I will be putting it through rigorous real-world tests across the MacoLabs product portfolio.

Final Thoughts

A software development workflow should never be a static dogma—it is a living, adapting system. In March, Gemini provided the initial breakthrough away from WordPress; in June, I separated Antigravity from Codex; and by September, sheer practical results pushed 90% of the work onto 5.6sol (medium-high).

The tools will keep shifting, but the underlying mission remains the same: shipping stable, high-value digital products with the least possible friction.