Tably Team
OpenAI’s GPT models are the best known on the market and, for many people, synonymous with artificial intelligence. At Tably Team we use them every day, but not for everything: they have a profile of their own that is worth understanding before giving them a seat on the team.
The current family
OpenAI maintains two lines that coexist:
- GPT (full version): the main model, with extended reasoning. It is the right choice for leadership, product and scope decisions.
- GPT mini: a smaller, much cheaper, fast version, surprisingly capable on narrowly scoped tasks. It is the default model Tably proposes for OpenAI because it performs very well relative to its cost.
There is also Codex CLI, OpenAI’s coding agent, which runs on your machine with your ChatGPT account and which Tably can use as a desktop builder.
What sets it apart
Conversation and synthesis
GPT is excellent at summarizing, rephrasing and explaining. In the Tably council, a “GPT” CEO is usually the one who best distills the debate into three clear decisions and communicates them without jargon.
Breadth of general knowledge
For a product role that has to reason about a specific industry (a clinic, an accounting firm, a clothing store), GPT brings very broad business context and proposes realistic flows.
Format and structure
It produces well-organized output (tables, lists, numbered criteria) with great consistency. For the Direct builder’s brief, when no Anthropic API is configured, it is a great second option.
Tools and ecosystem
The ecosystem around OpenAI is enormous: SDKs, integrations, documentation. For the Head of Engineering who has to propose libraries and services, that shows in the quality of the recommendations.
Where you should not use it
- Very long builds on a repository: Codex works well, but in our experience Claude Code finishes more projects without intervention. If you have both subscriptions, try both with the same brief and compare.
- Faithful replicas of a reference: GPT tends to “improve” the design on its own initiative. If you asked for a clone, Claude sticks to the blueprint more closely.
- When you need hard technical disagreement: GPT is diplomatic; for a CTO who raises uncomfortable objections, DeepSeek or Claude are usually more direct.
Which role to give it on your team
- CEO: our most frequent choice. It decides, prioritizes, delegates and communicates.
- Product: very good. User flows, acceptance criteria, business questions.
- Design: solid, with good general taste; for very strict design systems we prefer Claude.
- QA: the mini version reviews consistency quickly and cheaply.
- Builder: Codex CLI with your ChatGPT subscription, or GPT via API. It works; compare it with Claude Code in your own case.
How it connects in Tably
- OpenAI API key in “Models and API keys”. It enables GPT for any role and as a builder via API.
- Codex CLI: if it is installed and signed in on your machine, it appears as “Codex CLI · desktop” in the builder selector, without spending API tokens.
Costs
GPT mini is one of the cheapest models on the market for its level; the full version sits in the high range, similar to Opus. Tably estimates the council’s cost before starting and adds up the real tokens when it finishes. With Codex via subscription, the build does not spend API tokens.
Frequently asked questions
Can I use my ChatGPT account instead of an API key? For building, yes: through Codex CLI. For the council roles you need an API key.
GPT or Claude for the CEO? Both work; GPT communicates better, Claude is more conservative with scope. Our favorite combination is a GPT CEO and a Claude CTO: they really debate.
Is the mini model good for anything serious? For QA, summaries, classification and descriptions, without a doubt. For architecture, use the full version.
In short
GPT is the team model that thinks best about the business and communicates best. Give it leadership and product; for code, measure it against Claude in your own project and keep whichever works best for you.
