Which AI tool and model should you use when?
Our Claude Code and Codex workflow, with dated model examples and a practical way to choose quality, speed, and cost.
01Choose the workbench, then the model
Claude Code and Codex are working environments with tools, repository context, and permissions. The model is one component inside that environment. A model name alone does not tell you which browser, image tool, integrations, or account limits you have.
Nicholas’s starting workflow is shown below. The model roles are editorial starting points, not a cross-vendor benchmark or a claim that each named model is used on every engagement.
02Our starting tool split
Switch roles when the task or observed output justifies it. Give either tool the same approved requirements and design references.
| Work | Usual starting point | What remains a human responsibility |
|---|---|---|
| Product design | Claude /design | Define the user journey, critique options, and approve the interaction design. |
| Application implementation | Claude Code; sometimes Codex or both | Confirm architecture and inspect the behavior being added. |
| Code review | Codex, with permission to read the connected code | Reproduce important findings and decide what must change. |
| Images and logo exploration | Codex with its image-generation tools | Art direction, brand consistency, asset checks, and final suitability. |
| Landing pages and websites | Codex, using an approved brief and visual direction | Claims, content, responsive behavior, accessibility, and conversion flow. |
03Concrete model examples, checked October 28, 2026
Claude’s model aliases can resolve differently by provider. Confirm the actual model in your session rather than copying a version from an old article. OpenAI model availability and settings also depend on the product and account.
The following routing choices are our suggested starting points. Compare them on representative tasks before making them a project rule.
| Task shape | Model starting point | When to change |
|---|---|---|
| Routine Claude Code implementation | Sonnet family; Anthropic API example: Sonnet 5.5 | Try Opus for unresolved ambiguity or a harder architectural problem. |
| Difficult Claude Code investigation | Opus family; Anthropic API example: Opus 5.5 | Return to a lighter model once the work is well specified. |
| Focused, low-risk Claude tasks | Haiku family; Anthropic API example: Haiku 5.5 | Escalate if results repeatedly miss the acceptance criteria. |
| Codex implementation or review | GPT-6.1 Sol | Try GPT-6 Astra for difficult, high-consequence analysis. |
| Bounded OpenAI tasks | GPT-6 Luna | Use a stronger model when verification exposes a quality gap. |
| Visual analysis and writing | GPT-6.1 Terra | Evaluate on your real screenshots and deliverable. |
04Spend more reasoning where the decision needs it
A small copy edit and a permissions migration should not have the same budget. Start with a setting appropriate to the task, measure whether the output passes the checks, and increase effort when better analysis is worth the delay and usage.
Compare total work: retries, human corrections, and review time as well as token or subscription cost. More agents and more review rounds are not automatically an improvement.
05Image generation is its own capability
Using Codex as the workbench for a visual asset does not mean the selected coding model is also the image generator. Record the tool or image model used when that matters to the handoff. A generated logo direction also needs checks at small sizes and suitable final file formats.
Use Claude /design for the product interface workflow described here. A visual-generation task can be handled through Codex, then brought into the approved design system.