Step 1 · Describe the task

Choose the answer that best describes the task, not the answer that feels safest.

1How repeatable is the task?

Would a clear checklist work most of the time?

Very repeatableClear steps or a known format

Somewhat repeatableFamiliar, with a few judgment calls

New or open-endedNo reliable checklist yet

2How much judgment does the task require?

Is the main challenge execution, synthesis, or deciding between tradeoffs?

Very littleFollow instructions or transform material

SomeInterpret and make routine choices

A great dealWeigh ambiguous tradeoffs

3What happens if the answer is wrong?

Think about stakes and how easily a person can reverse the result.

Easy to reverseA quick correction is fine

Worth getting rightNormal professional consequences

ConsequentialCostly, sensitive, or hard to undo

4How many sources, constraints, or steps are involved?

Count competing inputs, requirements, and dependencies.

One or twoA focused request or source

SeveralNormal multi-step work

Many or conflictingEvidence must be reconciled

5How easy is it to verify the result?

Can a person quickly tell whether it is good enough?

Easy to checkClear facts, format, or quality bar

Needs reviewGood judgment is visible with context

Hard to knowSubtle errors may be missed

6What matters most for this run?

Volume and speed can justify a lighter option when the work is clear.

High volume or fast turnaroundKeep cost and latency down

BalancedNormal quality, speed, and cost

Quality over speedTime is available for a stronger pass

7Is this the final critical review?

Reserve this for a difficult final pass where a small reliability gain matters.

NoThis is working output or a normal decision

Almost finalImportant, but not a last-chance review

Yes, final and criticalA small miss would materially matter

Benchmark context

Cost changes the answer.

Workload fit comes first. The benchmark adds a second lens: some smaller-model, higher-effort combinations can deliver compelling intelligence per dollar.

Artificial Analysis chart comparing GPT-5.6 Sol, Terra, Luna, and GPT-5.5 intelligence index against weighted cost per benchmark task. — Source: Artificial Analysis Intelligence Index. Cost is the weighted average per Artificial Analysis benchmark task, not OpenAI pricing or a guarantee for real Codex workflows.

Advanced choices beyond the selector

Light, Medium, High, and Extra High tune the selected model. Use the lowest effort that produces the result you need. The selector reserves Light for tightly bounded Luna work and Extra High for unusually difficult or final reviews.

Max and Ultra solve different problems. Max gives one model more time for the hardest single task. Ultra brings in subagents for work that divides into meaningful parallel lanes. Most tasks need neither.

Sources

Guidance and benchmark evidence

OpenAI guidance determines workload fit. The third-party benchmark informs value alternatives. Test important workflows with representative work.

Choose for task fit first, then compare the value alternative. Escalate when ambiguity or the cost of being wrong outweighs the extra usage. · A free tool from Majestic Labs.

Keep thinking with us

Practical AI ideas, delivered where you already are.

Get occasional field notes on choosing models, building useful AI workflows, and making better decisions with the tools.

Keep me posted by email

No fixed schedule. We will write when there is something worth sharing.

Confirm once by email · Unsubscribe anytime.

Prefer a messaging app?

Telegram and WhatsApp are broadcast-only and carry the same posts. Pick the app you prefer.

Telegram WhatsApp