Which model should you use?
Every provider ships roughly three tiers. Matching the tier to the task is the single biggest lever on cost — and reaching for the most powerful model by reflex is the most common way to overspend. The rule: start balanced, escalate when the answer disappoints, drop to fast for anything you run in bulk.
The same three tiers, across providers
| Tier | Claude | ChatGPT | Gemini |
|---|---|---|---|
| Fast & cheap High-volume, low-judgment tasks | Haiku 4.5 | GPT-5.6 Luna | Gemini 3.7 Flash |
| Balanced Everyday drafting & synthesis | Sonnet 5 | GPT-5.6 Terra | Gemini 3.7 Flash |
| Max reasoning Hard thinking & agentic work | Opus 5 · Fable 5 | GPT-5.6 Sol | Gemini 3.1 Pro |
Claude splits "fast" and "balanced" into two models (Haiku and Sonnet); ChatGPT and Gemini fold both into one everyday model and reserve a separate tier for heavy reasoning.
Low-reasoning, high-volume
Summarising, extracting, classifying, reformatting — anything well-defined that you'd run many times. Give it a clear rubric, not an open question. This is the tier people forget exists, and running these jobs on a flagship model is just burning budget.
Example queries
Summarise these notes into the 5 key points. One line each, no preamble.
{paste your notes}Find every place the old product name appears across these files so I can rename it. List file and line only — no commentary.Classify each feedback line into exactly one theme:
[onboarding, pricing, performance, bug, feature-request, other].
Output "line | theme". No explanations.
{paste feedback}Your everyday default
Drafting, rewriting, synthesising a handful of inputs, fast back-and-forth iteration — roughly 80% of product work. Start here for almost everything and only escalate when the output genuinely falls short.
Example queries
I'm thinking of calling it "ProdPrompts" — help me brainstorm a sharper name. Give me 6 options across different angles, with a one-line rationale each, and your top pick.A competitor may already have a tool like the one I'm building. Double-check what they actually ship, then tell me — in a short table — how mine differs.Draft a release note for this change. Audience: existing customers. Tone: clear, no hype. Max 80 words. Lead with the user benefit.
Change: {describe the change}Hard thinking & agentic work
Strategy, trade-offs, deep analysis, and long-horizon agentic builds — multi-step tasks that run with autonomy and where a shallow answer is expensive. Worth the wait and the cost for the 10–20% of work that actually carries weight.
Example queries
Evaluate this product idea: why it could matter for my career, how I'd get real traction, and where it could fail. Be candid, give a recommendation, and ask me anything you need.
{describe the idea}Build the whole thing: scaffold the site, wire up the pages and data, and verify it actually runs before you hand it back. Tell me what you checked.Analyse how I've been prompting across our recent work and extract insights on when I get the best results — what's working, and where I could tighten.The 10-second cheat sheet
| Summarise, extract, classify, reformat at volume | Fast |
| Draft, rewrite, synthesise a handful of inputs | Balanced |
| Strategy, multi-step builds, deep analysis | Max reasoning |