The Two Dials, In Plain English
Model = how much it knows. A bigger model has seen more, recognizes more patterns, and catches things a smaller one walks right past.
Effort = how hard it tries. Same model, more thinking. It reads more before acting. It double-checks. It actually looks at your brand docs instead of guessing.
These are not the same thing, and this is the part almost everyone gets wrong. A small model on max effort will think very hard and still be wrong, because it doesn't know the thing. A big model on low effort will know the answer and still miss it, because it didn't bother to look.
Anthropic's own guidance boils it down to one question when a task goes badly:
"Did it not try hard enough, or did it not know enough?"
Didn't try hard enough (skipped your files, ignored half the brief, quit halfway) means raise the effort. Didn't know enough (had everything it needed, still confidently wrong) means raise the model. That one question replaces about 90% of the guesswork.
Claude And Codex, Side By Side
Same four tiers, different names. If you use one tool, this table tells you what the other one calls the same thing. Current as of August 2026.
| Tier | Claude Code | Codex |
|---|---|---|
| Cheap and fast | Haiku 4.5 | GPT-5.6 Luna |
| Everyday workhorse | Sonnet 5 | GPT-5.6 Terra |
| Heavy thinking | Opus 5 (the default) | GPT-5.6 Sol (the default) |
| Hardest, longest jobs | Fable 5 | Sol at Max or Ultra |
The effort levels:
- Claude Code: low, medium, high (the default), xhigh, max. Plus
ultracode, a session-only mode that thinks at xhigh and plans a workflow for every task. - Codex: low, medium (the default), high, extra high, max, ultra. Max gives one model more thinking on one hard problem. Ultra splits the job across helper agents.
The cheap tier should run hot.Luna is so cheap that there's almost no reason to run it on low. Crank it. A maxed-out Luna still costs less than a lazy Sol. Haiku 4.5 has no effort dial at all, so there's nothing to set. Just use it.
The scale is calibrated per model.High on Sonnet 5 is not the same amount of thinking as high on Opus 5. Don't copy settings between models and expect the same behavior.
30 Real Tasks, Mapped
Not developer tasks. The stuff you actually do to run and grow a business.
Grunt work (cheap model, effort cranked)
| # | Task | Claude | Codex |
|---|---|---|---|
| 1 | Pull names and emails out of a messy spreadsheet | Haiku 4.5 | Luna, max |
| 2 | Tag 500 leads by industry or deal size | Haiku 4.5 | Luna, max |
| 3 | Clean up a raw transcript into readable text | Haiku 4.5 | Luna, max |
| 4 | Summarize a 90-minute call into bullets | Haiku 4.5 | Luna, max |
| 5 | Reformat a product list into a clean CSV | Haiku 4.5 | Luna, max |
| 6 | Rename and sort a folder of 200 files | Haiku 4.5 | Luna, max |
| 7 | Write alt text for a batch of images | Haiku 4.5 | Luna, max |
| 8 | Pull every action item out of your notes | Haiku 4.5 | Luna, max |
Everyday production (workhorse model)
| # | Task | Claude | Codex |
|---|---|---|---|
| 9 | Turn an approved script into 5 social captions | Sonnet 5, medium | Terra, medium |
| 10 | Repurpose one post for LinkedIn, X, and IG | Sonnet 5, medium | Terra, medium |
| 11 | Write emails 2 through 5 after you approved email 1 | Sonnet 5, medium | Terra, medium |
| 12 | Turn a recorded process into a written SOP | Sonnet 5, medium | Terra, medium |
| 13 | Update the copy on a page you already built | Sonnet 5, medium | Terra, medium |
| 14 | Build a landing page from a spec you already wrote | Sonnet 5, high | Terra, medium |
| 15 | Fix a small thing that broke on your site | Sonnet 5, medium | Terra, medium |
| 16 | Connect a form to your CRM (simple automation) | Sonnet 5, high | Terra, high |
| 17 | Build a one-page report from a spreadsheet | Sonnet 5, high | Terra, high |
Money work (heavy model, being wrong costs you here)
| # | Task | Claude | Codex |
|---|---|---|---|
| 18 | Write a sales email to your list | Opus 5, high | Sol, high |
| 19 | Write a video script or a hook set | Opus 5, high | Sol, high |
| 20 | Research a topic across many sources | Opus 5, high | Sol, high |
| 21 | Competitor research and how you position against them | Opus 5, xhigh | Sol, extra high |
| 22 | Design a landing page from scratch (concept, copy, layout) | Fable 5, medium | Sol, extra high |
| 23 | Write a full sales page or VSL | Fable 5, medium | Sol, extra high |
| 24 | Name and price a new offer | Opus 5, xhigh | Sol, extra high |
| 25 | Design a multi-step automation across 3+ tools | Opus 5, xhigh | Sol, extra high |
| 26 | Figure out why an automation silently stopped working | Opus 5, xhigh | Sol, extra high |
| 27 | Look at your numbers and tell you what to change | Opus 5, xhigh | Sol, extra high |
| 28 | Build a real dashboard from your live data | Opus 5, max | Sol, max |
Rows 22 and 23 look like a typo until you know the trick. Fable 5 at medium effort still beats older models running flat out. When the output is the revenue engine, you buy the better model, not more thinking. Cheaper and better at the same time.
Big builds (heaviest model, leave it running)
| # | Task | Claude | Codex |
|---|---|---|---|
| 29 | Build a full app or SaaS MVP | Fable 5, xhigh | Sol, ultra |
| 30 | Move your whole business off one platform onto another | Fable 5, max | Sol, ultra |
Notice the shape. The top 17 rows are most of your week, and none of them need your most expensive setup. The heavy models show up where a bad answer costs you a launch, a client, or a month.
Four Questions Before You Pick
1. What does it cost me if this is wrong? A bad caption costs nothing, you see it instantly and rewrite it. A bad sales page costs you a launch. Cheap to catch means low effort. Expensive to catch means high.
2. Do I already know the answer?If you can describe exactly what you want, you don't need thinking. You need typing. Low to medium. If you're saying "figure out why this isn't working," you're buying thinking. Go high.
3. How many moving parts? One doc, one obvious change: small model. Five tools, unclear scope, decisions inside decisions: big model.
4. How many times will I run this? Once? Spend whatever. Two hundred times in a loop? Now effort is a budget line, not a quality choice.
Then follow the escalation rule: start one notch below where you think you need to be, and escalate on evidence, not on nerves. If medium actually failed, go high. If medium just feltrisky, you're paying a tax on your own anxiety.
Four Habits That Cost Real Money
Running max on everything.Max is prone to overthinking. It burns tokens circling a problem that high solved in a third of the time. It's a precision tool, not a default.
Confusing thinking with talking. Effort controls how much the model reasons internally, not how long the answer is. On Opus 5, cranking effort does not reliably make the response longer or shorter. If you want a short answer, ask for a short answer.
Never using the cheap tier.Luna and Haiku exist for a reason. Extraction, tagging, cleanup, summaries. If a task has one obvious right answer, you're lighting money on fire using anything bigger.
Planning cheap, building expensive. Backwards. The best pattern is the opposite: use a heavy model at high effort to make the plan, then drop to a cheaper model at medium to execute it. Claude Code has this built in as opusplan. Codex users do it manually with plan mode on Sol, then switch to Terra for the work.
The Actual Commands
Claude Code:
You can also set effortLevel in your settings file so it sticks, or set effort inside a skill so that job always runs at its own level. And if you just want one message to think harder without changing anything, put the word ultrathink anywhere in your prompt.
Codex:
Or set model_reasoning_effort in your config.toml to make it permanent, and plan_mode_reasoning_effort for planning.
The One-Line Version
Cheap fan-out, expensive judgment. Let the small models do the volume. Save the deep thinking for the decisions where being wrong actually costs you money.
That's the whole game.