Current operating model · Updated July 10, 2026
My AI stack got smaller as the general-purpose systems learned to use tools, work across files, and stay with a task longer. The useful question is no longer “which app can do this?” It is “which system should own this step, and what context does it need?”
The model is not the operating model. The operating model is the loop around it: source material, tools, boundaries, verification, and a durable place for the work to live.
The current loop
These are roles, not rigid tool boundaries. The point is to keep the decision thread intact while each system does the work it is best positioned to finish.
01 · Frame
ChatGPT
I use ChatGPT for operating models, product strategy, research synthesis, process architecture, and the conversations where the problem is still taking shape.
Best for: Ambiguity, competing frames, and decisions that need a clear throughline.
02 · Build
Claude + Codex
This is the execution layer: repositories, terminals, browsers, files, tests, and long-running work that has to end in a finished artifact rather than a persuasive answer.
Best for: Implementation, inspection, debugging, and verified multi-step delivery.
03 · Capture
PLAUD + source files
Ideas and conversations are captured before they are polished. The transcript or source artifact stays available so later synthesis can be checked against what was actually said.
Best for: Voice notes, meetings, early requirements, and evidence that should not be overwritten.
04 · Preserve
Project repos + curated knowledge stores
Important work lands in a durable home with decisions, source material, current state, and clear boundaries. The model changes. The context should survive it.
Best for: Long-running projects, repeatable workflows, and anything another agent may inherit.
GPT-5.6 routing
OpenAI’s durable tier names make the routing decision visible. Sol is the flagship. Terra balances capability, speed, and cost. Luna is the fastest and most efficient. In ChatGPT Work and Codex, the useful choice is the lane that can finish the task with the least total review debt.
Flagship reasoning
The highest-capability GPT-5.6 tier for complex professional work across coding, research, knowledge work, computer use, science, cybersecurity, and design.
Route here when
Balanced execution
The capability, speed, and cost balance in the family. This is the useful default for everyday professional work that still needs planning and tool use.
Route here when
Fast, efficient throughput
The fastest and lowest-cost GPT-5.6 tier. Its value is not “less important work.” It is bounded work where speed, volume, and easy verification matter most.
Route here when
Sol is rolling out in standard ChatGPT reasoning modes. Terra and Luna are available in ChatGPT Work, Codex, and the API rather than as standard chat model-picker options.
Claude’s lane
Claude remains where I build and do. Sonnet 5 strengthened the part of the product I already valued: sustained coding, browser and terminal use, tool calling, and the ability to carry messy work through the verification pass.
ChatGPT and Claude overlap more than they used to. I still get better results when I give them different jobs. ChatGPT holds the higher-altitude frame. Claude works closer to the artifact. Codex joins that execution layer when the repository and the release path are the center of the task.
The handoff rule
Do not hand off a polished answer. Hand off the source, the decision, the boundary, and the definition of done.
That is what lets another model continue the work without inventing the thread you forgot to preserve.
The agent test
A chat with instructions is useful. An agent is a repeatable system with a trigger, a process, tools, context, and explicit stopping rules.
What I retired
It proved that dense information could become interactive. It also froze a moment when the tool layer was fragmented and hypothetical use-case catalogs felt useful. The market moved. My own practice moved faster.
The code remains in the repository history. The public page now reflects the operating model I would actually defend.
Source shelf
Capability and availability claims were checked against the current GPT-5.6 announcement, ChatGPT availability guide, workspace-agent guide, and Claude Sonnet 5 release. The curation model was informed by how FutureTools separates recommendations from the full archive and how AI Daily Brief keeps current changes concise. The structure and point of view here are my own.