Most people using Claude or ChatGPT still do not know that there is now a dial controlling how hard the model thinks before it answers. It is not a prompt trick. It is a setting.
Anthropic shipped it with Claude Opus 5 on 24 July 2026 as an "effort" control. OpenAI shipped its version with GPT-5.6 on 9 July 2026 as a max reasoning setting plus an "ultra" mode. Two of the largest labs added the same class of control within two weeks of each other, which tells you something about where the leverage is.
If you have been trying to fix inconsistent outputs by rewriting prompts, this is the variable you have been missing.
What is an effort setting in an AI model?
An effort setting is a control that tells a model how much reasoning to spend on a request before it answers. High effort means more internal deliberation, better accuracy on hard tasks, more tokens burned and slower replies. Low effort means faster, cheaper, shallower answers. It is a cost-quality dial, not an on-off switch.
The important part is that effort is separate from model choice. You are no longer only deciding which model handles a task. You are deciding how much thinking that model is allowed to do.
That distinction matters because most disappointing AI output is not a model failure. It is a mismatch: a task that needed deliberation was answered at conversational speed.
Anthropic describes the Opus 5 effort setting as a way to optimise for intelligence at one end, or conserve tokens for faster and cheaper results at the other. OpenAI frames its equivalent as reasoning depth, with the highest tier able to split a task apart and run parallel subagents that coordinate with each other.
What did Anthropic and OpenAI actually ship in July 2026?
Both labs shipped reasoning controls attached to new flagship models. Anthropic released Claude Opus 5 on 24 July 2026 with an effort setting. OpenAI released GPT-5.6 on 9 July 2026 as three tiers with a max reasoning setting and an ultra mode. The controls are new, so most published prompting advice predates them.
Here are the specifics worth knowing, per each company's own announcements:
--- Claude Opus 5 (Anthropic, 24 July 2026): US$5 per million input tokens, US$25 per million output tokens. 1M-token context window, 128K max output. Introduces the effort setting. Anthropic reports it lands within 0.5% of Claude Fable 5's peak on CursorBench 3.2 at roughly half the cost per task.
--- GPT-5.6 (OpenAI, 9 July 2026): ships as three tiers instead of one model. Sol at US$5 input / US$30 output per million tokens, Terra at US$2.50 / US$15, Luna at US$1 / US$6.
--- Ultra mode: on the Sol tier, the model can break a task apart and run parallel subagents. OpenAI reports this lifts its score on a command-line coding benchmark from 88.8% to 91.9%.
--- Where you see it: in the API you set it explicitly. In the consumer apps it is partly handled for you, which is why the same question can produce noticeably different depth on different days.
Treat these numbers as version-specific. Model families are re-tiered and re-priced often, and a benchmark score attached to one release does not transfer to the next.
When should you turn effort up, and when should you leave it down?
Turn effort up when the task has a wrong answer that costs you something: analysis, planning, anything with constraints that interact. Leave it down when the task has no hidden trap: formatting, summarising, rewriting, classification. High effort on a simple task does not improve the output. It just costs more and takes longer.
A working routing rule, based on how these controls behave in practice:
--- Low effort / cheap tier: reformatting, tone edits, extracting fields from a document, tagging inbound messages, drafting variations of copy you will edit anyway.
--- Medium effort / balanced tier: first drafts of real deliverables, research summaries with a defined scope, turning meeting notes into a structured brief.
--- High effort / flagship tier: pricing and budget models, campaign plans where the channels trade off against each other, comparing vendors on criteria you have to defend, anything you will present to someone senior.
--- Ultra or parallel-subagent modes: multi-step work with tool use and retries. Overkill for a single question, and you will feel the latency.
The counterintuitive part: the biggest wins from high effort usually come from mid-sized ambiguous tasks, not from the hardest ones. A vague brief with three plausible interpretations is exactly where deliberation pays. A genuinely hard, well-specified task is often already handled well at medium effort.
There is a cost consequence too. Higher effort produces more reasoning tokens, and reasoning tokens are billed as output. On the numbers above, a high-effort task on GPT-5.6 Sol at US$30 per million output tokens is not something you want running on autopilot across a whole inbox.
How do you write a prompt that actually benefits from high effort?
High effort rewards prompts that give the model something to deliberate about: competing options, explicit constraints, and a named failure condition. It does nothing for a prompt that is just a request for text. If your prompt has one obvious answer, extra reasoning has nowhere to go.
The pattern that works: force a comparison, then force a discard. Give the model alternatives and a reason to eliminate them, and the extra thinking gets spent on judgement instead of padding.
Try this prompt:
You are reviewing a decision, not answering a question.
Task: [paste the actual decision you need to make]
Context: [paste the constraints, the budget, the deadline, and anything already ruled out]
Before you answer, do this internally:
1. List the three most plausible approaches.
2. Name the strongest objection to each one.
3. Discard any approach that fails its own objection.
Then output only:
--- Your recommendation, in one sentence.
--- The two assumptions it depends on.
--- What evidence would change your mind.
Do not show your intermediate reasoning. If the context is missing something you need, say what is missing instead of guessing.
Run that prompt twice on the same task, once at low effort and once at high. The low-effort version typically returns the most conventional option with confident framing. The high-effort version more often surfaces the objection you had not written down yet.
That last instruction matters as much as the effort setting. Reasoning models given an incomplete brief will reason their way to a plausible answer built on invented assumptions. Telling it to name the gap instead is what converts deliberation into something you can trust.
This is also why effort settings pair well with structured context. If you want the deeper background on assembling that context properly, our guide to context engineering covers the discipline that sits underneath this.
Where do effort and reasoning controls break down?
Effort controls fail in four predictable ways: they amplify bad context, they add latency that breaks interactive workflows, they make costs unpredictable when automated, and they cannot fix a task that was never well defined. None of these are bugs. They are the shape of the trade-off.
--- Garbage in, elaborate garbage out. High effort applied to a thin brief produces a longer, more confident, more internally consistent wrong answer. It is harder to spot than a short wrong answer, which makes it more dangerous.
--- Latency breaks the loop. If you are iterating conversationally, a high-effort reply that takes 40 seconds destroys the rhythm. Draft at low effort, then escalate the near-final version.
--- Automated cost drift. Effort set high inside a workflow that runs hundreds of times a day is the fastest way to a surprising bill, because reasoning tokens bill as output tokens.
--- It is not a factuality fix. More reasoning improves consistency and structure. It does not give the model information it never had. If the answer depends on your internal data, effort will not retrieve it.
--- Benchmarks do not transfer to your work. A reported jump from 88.8% to 91.9% on a coding benchmark says nothing about whether your quarterly plan comes out better. Test on your own tasks.
How do you test this on your own work in the next 20 minutes?
Pick one recurring task you already run through AI, run it three times at different effort levels with identical input, and compare the outputs side by side. Twenty minutes is enough to find out whether your default setting has been wrong. Most people discover they have been overpaying on simple tasks and under-thinking the important ones.
The test, step by step:
--- Step 1. Choose a task you do at least weekly and where you can judge quality yourself. A campaign brief, a vendor comparison, a client email that needs to land correctly.
--- Step 2. Write the input once and freeze it. Same words, no edits between runs, or the comparison is worthless.
--- Step 3. Run it at the lowest effort available, then medium, then the highest. Save all three.
--- Step 4. Score each output on one thing only: how much editing it needs before you would send it.
--- Step 5. Note the time each run took. If the highest-effort version needs the same amount of editing as the middle one, you have found your default, and you have just cut your cost per task.
Do this for your three most common AI tasks and you will have a routing policy of your own. That is worth more than any generic recommendation, because it is calibrated to your standards rather than a benchmark.
If you are already building repeatable systems rather than one-off prompts, the same testing discipline applies when you package a workflow as a reusable agent skill.
The takeaway
Prompt quality got most of the attention for three years. Effort is the variable that quietly arrived in July 2026 and changed the shape of the problem: the same prompt, the same model, two different answers, depending on a setting most people never touch.
The skill is not "always turn it up". It is knowing which of your tasks deserve deliberation and which just need speed, and having tested it rather than guessed. That is a twenty-minute experiment, not a course.
We know AI's cold edges. We know your real challenges. 28 years with UD, turning technology into a partnership with warmth.
You can verify the release details for yourself on Anthropic's newsroom, and always check the current model card before relying on a benchmark number.
Reviewed by the UD AI team.
Find Out Where Your AI Skills Actually Sit
Knowing about effort settings is one thing. Knowing which of your daily tasks are being handled at the wrong depth is another.
Take UD's AI IQ Test to benchmark how you are actually using AI today, then see where the gaps are.
And when you are ready to turn techniques into a workflow that runs reliably every time, we'll walk you through every step, from tool setup to workflow design and deployment.
