Published: September 7, 2026
Working with GPT
My work does not always arrive at a comfortable pace. There are periods when tickets pile up, several projects need attention at the same time, and clients are waiting for changes that cannot be delayed. I still need to understand every request and be responsible for the result, but doing every implementation step manually can turn a busy week into a backlog.
Since August 2026, Codex and the GPT-5.6 model family have become an important part of how I handle that pressure. I use them for client work and internal company projects, especially when demand is high. The biggest change is not that an AI writes code for me. It is that I now have a repeatable workflow for moving from a ticket to a reviewed plan and then to a verified implementation much faster.
Finding a workflow after Copilot
I did not move directly from GitHub Copilot to one permanent replacement. After Copilot changed how its usage was measured, I experimented with free models and different coding-agent tools. I used OpenCode with models such as DeepSeek-V4-Flash and Big Pickle. I also used 9Router to map models to agents, create combinations, and route work through different providers.
Those experiments helped me understand that the model is only one part of the experience. The workflow around it matters just as much: how the agent explores a project, how it explains its plan, where I can review its assumptions, and how it verifies the result.
I sometimes used Google Antigravity as well, and I especially liked its approach to implementation plans and review checkpoints. Being able to inspect the workflow at each stage made the agent easier to trust. It reinforced a habit that I still use now: planning and implementation should be separate decisions, with a human review between them.
Plan, review, accept, and implement
My regular Codex workflow is simple, but each step has a clear purpose:
- Plan with GPT-5.6 Sol at high reasoning, so the agent can inspect the project, understand the request, identify affected areas, and propose an implementation before changing anything.
- Review the plan myself, checking its assumptions, scope, technical direction, edge cases, and whether it actually solves the ticket.
- Accept the plan only when it looks right, or continue refining it when something is incomplete or too broad.
- Implement with Sol at low reasoning for the regular execution work, where the direction is already clear and speed matters.
- Use GPT-5.6 Terra at high reasoning when appropriate, depending on the complexity and shape of the task.
- Review and verify the result, because generated code still needs testing, judgment, and responsibility before it is complete.
I do not treat those model settings as rules for every project. Some tasks are small enough to solve directly, while others need more exploration or another review cycle. Sol high for planning and Sol low for implementation is simply the combination I use most often, with Terra high as a useful alternative when the work benefits from it.
From two days to one or two hours
This workflow has changed my effective working hours in a way I can feel. A slightly complex task that might previously have taken me around two days can sometimes be solved in one or two hours. That does not mean every two-day task is suddenly easy, and it is not a universal productivity percentage. It is a concrete result I have seen when the requirements are clear, the existing project is understandable, and the agent can work from a good plan.
The time saved is valuable because it gives me more room to handle urgent client requests without leaving other tickets untouched. It also lets me spend more attention on the decisions that matter: whether a feature fits the business need, whether the implementation is safe, and whether the result will remain maintainable after the immediate deadline.
I use this approach across different stacks. Codex helps me build custom PHP scripts, work with Odoo projects where much of the code is Python, develop Next.js applications, and handle broader full-stack tasks for my internal company work. Moving between those environments is easier when the agent can first study the repository and give me a plan that I can challenge before it starts implementing.
A trial that became part of my work
I started this current phase in August 2026 through a ChatGPT Pro trial. GoPay placed an Rp1 verification charge and returned it, so the trial was effectively free for me. That small opportunity gave me enough time to test the tools against real work instead of judging them from a few isolated prompts.
The coding limits have also felt generous in my experience. During a demanding project session, I can use around 80% of a five-hour window while the weekly meter decreases only slightly. I have not reached the full limit so far.
I was also given three earned rate-limit resets at unexpected moments. They appeared before I had exhausted my allowance, sometimes while the weekly usage meter was somewhere around 20–70%. I do not know why those resets were granted, and I do not assume every Pro account will receive the same thing. I only know that seeing more room become available while I was actively working made me genuinely happy and grateful.
More than a coding tool
The Pro experience has been useful beyond implementation. Longer conversations feel more practical because I can keep more context around a problem. I use ChatGPT Work for reports, deeper research, and other tasks that need more than a quick answer. Image generation has also been impressive, especially when I need to turn an idea into something visual without interrupting the rest of my workflow.
Together, these tools give me a connected way to work. A request can begin as research, become a plan, move into a software change, and finish as a report or visual output. I can keep my attention on the purpose of the work while the agent helps with the effort required to move between those stages.
What makes the workflow valuable is still the review loop. I do not accept a plan because it came from Sol, and I do not ship an implementation because an agent completed it quickly. I accept the plan when it makes sense, and I ship the work when the result has been reviewed and verified. Codex and the GPT-5.6 family give me leverage, but the responsibility remains mine.
So far, this has been a much better agent experience for the way I work. It helps me respond when demand is high, finish complex tasks in a fraction of the time, and use more of my day for judgment instead of repetitive implementation. The speed is exciting, but the real improvement is being able to serve my clients and support my company without letting the quality of the work fall behind the urgency.