Model vs. effort

Extended lesson · 14 slides · about 7 minutes

1 of 14

What changes when you raise effort?
Effort is not a new model. It changes how much reasoning and compute the selected model spends on the current response.
More planning before answering: The model spends more time breaking down the problem before it responds.
More checking of steps and assumptions: It double-checks its work and considers more possibilities.
Better handling of multi-step or ambiguous tasks: It can think through more complex problems with deeper reasoning.
Usually slower response and more compute: Higher effort takes longer and uses more resources.
Still not a guarantee of correctness: More effort can improve quality, but it doesn't make the answer always right.
Effort is a depth knob, not a magic intelligence upgrade.

The general effort ladder
Same task. Different effort level. Big difference in cost and result.
Low / fast: Routine rewrites, simple extraction, basic Q&A. Fastest and lightest.
Medium: Everyday analysis, drafting, spreadsheet help, ordinary coding. Balanced depth and speed.
High: Multi-step planning, harder debugging, deeper document analysis, decisions with more tradeoffs.
Extra high: Long or difficult technical work, agentic loops, complex synthesis. Slower and more resource intensive.
Max / deepest: Only where a product offers it. Hardest correctness-sensitive tasks where the extra time and usage are justified.
Vendors use different names. The idea is the same: more effort = more reasoning budget, not a different brain.
Use the lowest effort that reliably clears your quality bar.

ChatGPT: model vs. effort
The thinking level can change without changing the base model.
GPT-5.6 Sol powers multiple effort levels: On eligible paid ChatGPT plans, GPT-5.6 Sol powers Instant, Medium, High, and Extra High thinking levels.
Four effort levels
Instant: Fastest everyday responses.
Medium: Standard thinking for most tasks.
High: Extended thinking for harder problems.
Extra High: Highest GPT-5.6 Sol effort where available.
Pro is different: GPT-5.6 Sol Pro is a Pro-model option, and GPT-6 Pro is available on eligible plans.
Free and Go use Luna: Free and Go plans use GPT-5.6 Luna as the default. Think uses Luna for harder questions.
Availability depends on your plan: Which models and effort levels you can use depends on your plan and workspace settings.
In ChatGPT, raising effort may keep the same model. Choosing Pro can change the model.

ChatGPT: which effort should you use?
Start fast. Escalate only when the task earns it.
Instant: Quick Q&A, rewrites, summaries, simple extraction, ordinary email drafts.
Medium: Everyday analysis, spreadsheet help, normal coding, structured planning, business writing.
High: Harder debugging, architecture choices, deep document analysis, multi-step planning, complex comparisons.
Extra high: Difficult technical synthesis, long chains of dependencies, high-stakes internal analysis where extra checking is valuable.
Pro: Use a Pro model for the hardest or longest-running work when model capability itself matters, not just more effort.
Don't run every errand with a race car. Match the setting to the job.

Claude: which effort should you use?
High is the default sweet spot. Go lower for routine work, higher for the hard stuff.
Low: Routine rewrites, extraction, simple Q&A, lightweight drafting. Best when speed and usage matter more than depth.
Medium: Everyday writing, summaries, analysis, and ordinary work where some extra reasoning helps.
High: Anthropic's recommended overall balance for quality and speed. Good for deeper analysis, planning, coding, and complex document work.
Extra high (xhigh): Designed for long-running coding and agentic tasks on supported newer models. More reasoning than High without always going all the way to Max.
Max: The most thorough option for the hardest, correctness-sensitive work. Expect the longest response time and highest usage.
Start with High for serious work. Drop down for routine tasks. Use xhigh or Max only when the problem earns it.

Claude vs. ChatGPT: don't map the labels 1:1
Different names. Same idea. Understand what they actually mean.
ChatGPT
Effort levels on GPT-5.6 Sol (where available)
Instant: Fastest responses. Lower reasoning effort.
Medium: Good balance of speed and depth.
High: More reasoning for complex tasks.
Extra High: Deep reasoning for the hardest problems.
Pro: A separate Pro-model option that can change the underlying model.
The names differ. The idea is the same:
Model = which engine.
Effort = how much reasoning.
Claude
Effort levels (on supported models)
Low: Faster responses. Lower reasoning.
Medium: Balanced depth and quality.
High: More reasoning for complex tasks.
Extra High (xhigh): Deeper reasoning for challenging work.
Max: Maximum reasoning and most thorough responses.
Thinking: A separate control on models where it can be changed. Some newer models keep thinking on.
Compare concepts, not labels. "High" in one product is not a guaranteed match for "High" in another.

When should you raise effort?
Use higher effort when the task is more complex, risky, or uncertain
The task has multiple dependent steps: When success depends on doing several things in the right order, higher effort helps plan and keep everything consistent.
There are meaningful tradeoffs or ambiguity: If there isn't a single clear answer and you need to weigh options, higher effort leads to more balanced and thoughtful results.
The answer must synthesize several documents or sources: When you need to combine information from multiple files, tools, or data sources, use higher effort to get a more accurate and complete answer.
Debugging or analysis has already failed once or twice: If earlier attempts didn't work, higher effort can help the AI take a fresh, deeper look and try a more thorough approach.
The cost of a wrong answer is materially higher than the cost of waiting: When mistakes could be expensive, risky, or hard to undo, take the time for a more careful response.
Keep effort lower when...
The task is routine, deterministic, easily verified, or time-sensitive.
Raise effort for complexity or consequence, not just because the task feels important.

What more effort cannot fix
Some AI problems are setup issues, not effort issues.
Missing or wrong data: If the source data is missing, incomplete, or incorrect, more effort won't make it accurate.
Vague goals or contradictory instructions: Unclear or conflicting instructions lead to confused results, no matter how much effort you put in.
Lack of required tool or system access: If the AI doesn't have access to the right tools, systems, or permissions, it can't get the job done.
An unsuitable base model: The wrong model for the task will struggle, no matter how much you prompt or refine the input.
Outdated information: If the AI needs current or live data, older information will produce outdated or incorrect results.
Missing human approval: For important or high-risk actions, AI still needs human review and approval before proceeding.
If the inputs, access, or model are wrong, turning up effort just makes the wrong setup work harder.

The tradeoff: cost, speed & quality
More effort can help, but the returns are not always linear.
Increase effort
Speed
Cost
Quality
Latency: Higher effort usually takes longer because the model spends more time reasoning, checking, or planning.
Usage: Higher effort can consume more compute and, in products like Claude, more tokens/usage. Subscription products may abstract some of that cost.
Quality: Hard tasks often improve with more effort, but easy tasks may barely improve at all.
Model cost: A stronger base model can cost more even before effort is increased. Model choice and effort are separate cost levers.
Business value: The right question is not "Which setting is smartest?" It is "Which model + effort reliably completes this task at an acceptable cost and speed?"
Measure cost per successful task, not just tokens, messages, or response time.

Small business takeaways: model vs. effort
Match the right model and effort to the work, so you get great results without wasting money.
Routine work: Use a fast/lower-effort setting for rewrites, summaries, extraction, ordinary email, and simple Q&A.
Serious knowledge work: Use Medium/High-type effort for analysis, planning, spreadsheet work, coding, and document review.
Hard or expensive-to-get-wrong work: Use the strongest appropriate model plus higher effort, then require human review for consequential decisions.
Current facts or business data: Give the AI the right files, search, databases, or connected tools. Effort cannot invent missing evidence.
Operating rule: Measure cost per successful task and use the lowest model+effort combination that reliably clears your quality bar.
Don't buy 'Maximum AI.' Buy the cheapest setup that reliably does the job. Model chooses the capability. Effort chooses the depth. Your workflow decides the value.

Lesson done.

Up next

What is model routing?

Where does this fit your business? Get your AI plan

Swipe to see each slide.

The short answer

The model is which AI engine you choose, with its own strengths, limits and price, while effort is how much reasoning the chosen model spends on one response. Use the lowest setting that reliably does the job, because higher effort is slower and uses more computing, and it cannot fix bad data or a vague request.

Three things to remember

01Changing the model changes what is possible; raising effort changes how thoroughly the same model works.
02Use low effort for routine work, and higher effort for multi-step, unclear or costly-to-get-wrong tasks.
03More effort cannot fix missing data, vague instructions or missing access, and important actions still need human review.

Every slide, in words

01
Model vs. effort
Two different dials that people keep mixing up.
Model
Which AI engine or capability profile you chose.
ChatGPT: GPT-5.6 Sol
ChatGPT: GPT-5.6 Luna
ChatGPT: GPT-5.6 Sol Pro or GPT-6 Pro
Claude (Opus, Sonnet, etc.)
Different strengths and capabilities
Different limits (e.g., context, modalities)
Different features available
Different cost and usage profile
Effort
How much reasoning and compute it spends on this response.
Low: Faster, lighter work
Medium: Good balance for everyday tasks
High: More thorough and detailed
Extra High / Max: Deepest reasoning (when available)
Same model, different effort levels
Spends more time exploring, checking, and planning
Uses more compute, tokens, and time
Often more thorough on complex tasks
Pick the brain first. Then choose how hard it should work.
02
What changes when you change the model?
The model is the underlying capability profile, not just a speed setting.
Reasoning & problem-solving: Stronger models can handle more complex, multi-step work and follow detailed instructions.
Coding, writing & domain skill: Models differ in strengths: some are better at code, others at writing, analysis, or specific industries like legal, healthcare, or finance.
Context & modalities: Supported input/output types (text, images, audio, video) and context window limits can be very different between models.
Tool/feature compatibility: Some models support different product features or API capabilities, like web search, file analysis, or function calling.
Price & speed profile: Stronger models may cost more per use and can respond more slowly, while smaller models are faster and more affordable.
A different model can change what is possible. Effort changes how hard that model tries.
03
What changes when you raise effort?
Effort is not a new model. It changes how much reasoning and compute the selected model spends on the current response.
More planning before answering: The model spends more time breaking down the problem before it responds.
More checking of steps and assumptions: It double-checks its work and considers more possibilities.
Better handling of multi-step or ambiguous tasks: It can think through more complex problems with deeper reasoning.
Usually slower response and more compute: Higher effort takes longer and uses more resources.
Still not a guarantee of correctness: More effort can improve quality, but it doesn't make the answer always right.
Effort is a depth knob, not a magic intelligence upgrade.
04
The four levers people confuse
They work together, but they solve different problems.
Model: Chooses the capability profile or engine.
Prompt: Tells it what job to do, constraints, and the desired output.
Context + data: Gives it the facts, files, history, examples, and evidence it can use.
Effort: Controls how much reasoning or computation the selected model spends on the response.
Bad data is not fixed by more effort. A vague prompt is not fixed by a bigger model.
05
The general effort ladder
Same task. Different effort level. Big difference in cost and result.
Low / fast: Routine rewrites, simple extraction, basic Q&A. Fastest and lightest.
Medium: Everyday analysis, drafting, spreadsheet help, ordinary coding. Balanced depth and speed.
High: Multi-step planning, harder debugging, deeper document analysis, decisions with more tradeoffs.
Extra high: Long or difficult technical work, agentic loops, complex synthesis. Slower and more resource intensive.
Max / deepest: Only where a product offers it. Hardest correctness-sensitive tasks where the extra time and usage are justified.
Vendors use different names. The idea is the same: more effort = more reasoning budget, not a different brain.
Use the lowest effort that reliably clears your quality bar.
06
ChatGPT: model vs. effort
The thinking level can change without changing the base model.
GPT-5.6 Sol powers multiple effort levels: On eligible paid ChatGPT plans, GPT-5.6 Sol powers Instant, Medium, High, and Extra High thinking levels.
Four effort levels
Instant: Fastest everyday responses.
Medium: Standard thinking for most tasks.
High: Extended thinking for harder problems.
Extra High: Highest GPT-5.6 Sol effort where available.
Pro is different: GPT-5.6 Sol Pro is a Pro-model option, and GPT-6 Pro is available on eligible plans.
Free and Go use Luna: Free and Go plans use GPT-5.6 Luna as the default. Think uses Luna for harder questions.
Availability depends on your plan: Which models and effort levels you can use depends on your plan and workspace settings.
In ChatGPT, raising effort may keep the same model. Choosing Pro can change the model.
07
ChatGPT: which effort should you use?
Start fast. Escalate only when the task earns it.
Instant: Quick Q&A, rewrites, summaries, simple extraction, ordinary email drafts.
Medium: Everyday analysis, spreadsheet help, normal coding, structured planning, business writing.
High: Harder debugging, architecture choices, deep document analysis, multi-step planning, complex comparisons.
Extra high: Difficult technical synthesis, long chains of dependencies, high-stakes internal analysis where extra checking is valuable.
Pro: Use a Pro model for the hardest or longest-running work when model capability itself matters, not just more effort.
Don't run every errand with a race car. Match the setting to the job.
08
Claude: model, effort & thinking
Three controls that are related, but not the same.
Model: Choose which Claude model you are using (for example, Opus, Sonnet, Fable where available). Models differ in capability, speed, and usage profile.
Effort: Supported models offer Low, Medium, High, Extra High (xhigh), and Max. Higher effort generally means more thorough responses, more time, and more token/usage consumption.
Thinking: A separate toggle that lets Claude spend more time reasoning before answering. On some newer Claude models, thinking is always on.
Availability: Exact models and effort levels depend on the model, plan, product, and your organization's settings.
Model: Opus / Fable
Effort: Low / Max
Thinking: Off / On
In Claude, changing effort does not automatically mean changing the model.
09
Claude: which effort should you use?
High is the default sweet spot. Go lower for routine work, higher for the hard stuff.
Low: Routine rewrites, extraction, simple Q&A, lightweight drafting. Best when speed and usage matter more than depth.
Medium: Everyday writing, summaries, analysis, and ordinary work where some extra reasoning helps.
High: Anthropic's recommended overall balance for quality and speed. Good for deeper analysis, planning, coding, and complex document work.
Extra high (xhigh): Designed for long-running coding and agentic tasks on supported newer models. More reasoning than High without always going all the way to Max.
Max: The most thorough option for the hardest, correctness-sensitive work. Expect the longest response time and highest usage.
Start with High for serious work. Drop down for routine tasks. Use xhigh or Max only when the problem earns it.
10
Claude vs. ChatGPT: don't map the labels 1:1
Different names. Same idea. Understand what they actually mean.
ChatGPT
Effort levels on GPT-5.6 Sol (where available)
Instant: Fastest responses. Lower reasoning effort.
Medium: Good balance of speed and depth.
High: More reasoning for complex tasks.
Extra High: Deep reasoning for the hardest problems.
Pro: A separate Pro-model option that can change the underlying model.
The names differ. The idea is the same:
Model = which engine.
Effort = how much reasoning.
Claude
Effort levels (on supported models)
Low: Faster responses. Lower reasoning.
Medium: Balanced depth and quality.
High: More reasoning for complex tasks.
Extra High (xhigh): Deeper reasoning for challenging work.
Max: Maximum reasoning and most thorough responses.
Thinking: A separate control on models where it can be changed. Some newer models keep thinking on.
Compare concepts, not labels. "High" in one product is not a guaranteed match for "High" in another.
11
When should you raise effort?
Use higher effort when the task is more complex, risky, or uncertain
The task has multiple dependent steps: When success depends on doing several things in the right order, higher effort helps plan and keep everything consistent.
There are meaningful tradeoffs or ambiguity: If there isn't a single clear answer and you need to weigh options, higher effort leads to more balanced and thoughtful results.
The answer must synthesize several documents or sources: When you need to combine information from multiple files, tools, or data sources, use higher effort to get a more accurate and complete answer.
Debugging or analysis has already failed once or twice: If earlier attempts didn't work, higher effort can help the AI take a fresh, deeper look and try a more thorough approach.
The cost of a wrong answer is materially higher than the cost of waiting: When mistakes could be expensive, risky, or hard to undo, take the time for a more careful response.
Keep effort lower when...
The task is routine, deterministic, easily verified, or time-sensitive.
Raise effort for complexity or consequence, not just because the task feels important.
12
What more effort cannot fix
Some AI problems are setup issues, not effort issues.
Missing or wrong data: If the source data is missing, incomplete, or incorrect, more effort won't make it accurate.
Vague goals or contradictory instructions: Unclear or conflicting instructions lead to confused results, no matter how much effort you put in.
Lack of required tool or system access: If the AI doesn't have access to the right tools, systems, or permissions, it can't get the job done.
An unsuitable base model: The wrong model for the task will struggle, no matter how much you prompt or refine the input.
Outdated information: If the AI needs current or live data, older information will produce outdated or incorrect results.
Missing human approval: For important or high-risk actions, AI still needs human review and approval before proceeding.
If the inputs, access, or model are wrong, turning up effort just makes the wrong setup work harder.
13
The tradeoff: cost, speed & quality
More effort can help, but the returns are not always linear.
Increase effort
Speed
Cost
Quality
Latency: Higher effort usually takes longer because the model spends more time reasoning, checking, or planning.
Usage: Higher effort can consume more compute and, in products like Claude, more tokens/usage. Subscription products may abstract some of that cost.
Quality: Hard tasks often improve with more effort, but easy tasks may barely improve at all.
Model cost: A stronger base model can cost more even before effort is increased. Model choice and effort are separate cost levers.
Business value: The right question is not "Which setting is smartest?" It is "Which model + effort reliably completes this task at an acceptable cost and speed?"
Measure cost per successful task, not just tokens, messages, or response time.
14
Small business takeaways: model vs. effort
Match the right model and effort to the work, so you get great results without wasting money.
Routine work: Use a fast/lower-effort setting for rewrites, summaries, extraction, ordinary email, and simple Q&A.
Serious knowledge work: Use Medium/High-type effort for analysis, planning, spreadsheet work, coding, and document review.
Hard or expensive-to-get-wrong work: Use the strongest appropriate model plus higher effort, then require human review for consequential decisions.
Current facts or business data: Give the AI the right files, search, databases, or connected tools. Effort cannot invent missing evidence.
Operating rule: Measure cost per successful task and use the lowest model+effort combination that reliably clears your quality bar.
Don't buy 'Maximum AI.' Buy the cheapest setup that reliably does the job. Model chooses the capability. Effort chooses the depth. Your workflow decides the value.

More in part 8

Next step

Where does this fit your business?

Six quick questions about your business, then your AI plan: what to start with, what a person still checks, and the lessons that go with it. Or bring it straight to a free thirty-minute call.

Get your AI plan Book a free call