What is up people!
If you use ChatGPT Plus or Pro, there is a small setting worth checking before your next difficult request. OpenAI changed how the product decides when to spend more time thinking, and adding “think harder” to your prompt will not change that setting for you.
That is easy to miss because the conversation still looks like a conversation. You can ask for a dinner suggestion, a cleaner paragraph, or a project plan full of competing deadlines in the same box. The request becomes more demanding without the interface necessarily making the change you expected.
The useful response is to get more deliberate about where you need extra reasoning. Some work deserves it. Some work needs a missing fact, a clearer requirement, or a simple calculation instead. Learning to tell those cases apart will help more than adding increasingly stern instructions to every prompt.
📌 TL;DR
Check the control → ChatGPT Plus and Pro no longer automatically move from Instant to deeper thinking for difficult requests. Choose an available thinking level when the work calls for it.
Diagnose the failure → Extra reasoning may help connect the facts you supplied. It cannot establish a fact you left out, and it should not quietly change a requirement to make a plan work.
Check the result → A quick, natural reply tells you little about whether the underlying work is correct or complete. This becomes especially important as voice systems talk while doing other work.
Change the setting when the facts have to fit together
In its September 14 release notes, OpenAI says it is retiring automatic switching from Instant to Thinking for Plus and Pro users globally. Those users can still select a deeper thinking option themselves. The change concerns how the product routes a request; it does not establish that the underlying models became less capable.
OpenAI’s current thinking guide makes the practical detail explicit: asking ChatGPT to “think harder” or “think deeply” does not automatically select a different thinking level. On Plus, Medium and High are available. The available choices differ by plan, and managed workspaces have their own controls, so check what your account actually shows.
A request to be careful can still improve the instructions. “Check whether any dates conflict” tells the assistant what to examine. But the words in the prompt and the selected amount of thinking are separate controls. You can ask a useful question while leaving the product on Instant.
There is a good reason to keep a fast option. If you need three alternatives to an awkward sentence, the quality is usually easy to judge as you read. You can reject the wording immediately, ask for another version, and get on with the document. A long wait would need to buy something you value.
Compare that with asking whether an exhibition can be ready before the doors open. The framed pictures arrive at a particular time. Hanging takes a fixed amount of time, photographs must happen afterward, and the opening is already advertised. Every individual sentence can be clear while the combined plan is impossible.
The second task requires the assistant to preserve relationships between facts. Moving photography earlier is only a solution if the pictures are already on the wall. Adding another volunteer only helps if the hanging time can change. A tidy timetable can conceal either mistake, which makes a first glance a poor test.
That is a sensible trigger for trying an explicit thinking level: the answer depends on several conditions being true together. You might be comparing proposals with different exclusions, checking a schedule, or finding the consequences of changing one step in a process. Choose the effort because of the work the answer must do.
Even then, the setting is an opportunity for a better answer, not a certificate attached to it. If a response is wrong, you still need to identify the mistake. Otherwise, you can spend longer waiting for a more elaborate version of the same misunderstanding.
Find the mistake before you ask for more thinking
Take a deliberately simple, fictional exhibition plan. The framed pictures arrive at 10 a.m. Hanging them takes 90 minutes and cannot begin before delivery. Photographing the finished display takes another 30 minutes and must happen after all the hanging. Both jobs must be finished before the doors open at 11:30.
Those requirements cannot all hold. Hanging finishes at 11:30, and photography finishes at noon. The plan is half an hour short even before allowing for a late van or a missing screw. That conclusion follows from the supplied times; no knowledge of the art world is required.
Suppose an assistant returns a polished schedule with photography beginning at 11. The problem is specific: it has overlapped two jobs you said must happen in sequence. Asking for a more enthusiastic plan will not help. Asking it to check each start time against the previous task’s finish gives it a concrete error to repair.
If that kind of missed dependency keeps happening, trying a higher thinking level is reasonable. You have a known failure to test. Keep the same facts and requirements, then inspect whether the new answer finds the conflict without changing them.
Now remove the delivery time from the request. You still know that hanging takes 90 minutes and photography takes 30, but the assistant no longer has enough information to say whether the morning works. A careful answer should identify the missing arrival time or give a condition: delivery must happen by 9:30 to finish at 11:30, with no spare time.
Extra thought cannot tell it when the van is booked. It might propose an earlier delivery, but that is an option to investigate. It becomes a fact only when somebody confirms it. This is why “use a smarter model” can be an expensive answer to an information problem.
There is another version of the mistake. Perhaps the assistant notices the clash and changes hanging from 90 minutes to 60. The arithmetic now works, but it has solved a different problem. If 90 minutes was a firm requirement, the response should say which condition needs to change and leave that decision with you.
A useful revision would be: “The current plan finishes at noon. To keep the 11:30 opening, we would need delivery by 9:30, a confirmed shorter installation, or permission to photograph afterward.” You can then check which option is possible. The assistant has made the decision clearer without pretending it has already been made.
Notice how the right next step changes across these examples. A missed connection calls for checking the reasoning. A missing arrival time calls for information. An impossible combination calls for changing a real constraint, with the relevant person’s agreement. All three can produce an answer that looks wrong, but they should not produce the same follow-up request.
You do not need to label every conversation this carefully. Use the distinction when you catch yourself asking for a third rewrite and the same underlying problem survives. Before sending another prompt, name one thing the answer failed to preserve. That often reveals whether the assistant needs more effort, better information, or a decision only you can make.
A smooth conversation still needs a correct result
The interface is becoming less helpful as a clue to how much work is happening underneath it. Google’s Gemini 3.8 Live announcement, published September 15 and updated September 17, describes a version called Extended Thinking that can reason while speaking. Google also says these models can carry out tool calls in the background while the conversation continues.
That is Google’s description, not a hands-on result from this newsletter. But the design has an immediate implication: a spoken acknowledgment can arrive before the task is finished. An assistant saying it is checking the delivery booking may still be waiting for the booking system.
The reverse inference is weak too. A long pause does not show that a schedule was checked correctly. You need the resulting times, the conflict it found, or the confirmed booking. How the conversation feels and what the assistant accomplished deserve separate attention.
There is a useful research example behind this distinction. Google cites EVA-Bench, a ServiceNow research framework published in May. Its researchers evaluated 12 voice-agent systems using simulated conversations across 213 scenarios in airline service, healthcare HR, and IT support. They measured accuracy separately from conversational experience.
For task completion, the framework checks whether the simulated system’s final records match the expected result. Other measures examine whether the agent follows the supplied information and speaks details correctly. The experience measures cover things such as conversational progress and turn-taking. A system can therefore be evaluated for making the right change as well as for being pleasant to talk to.
This was a benchmark using simulated callers, not a study of people doing their everyday work. Its older model results should not be treated as a verdict on this week’s Gemini release. The useful lesson is the measurement design: checking the conversation alone leaves part of the job unchecked.
Back at the fictional exhibition, “I’ve moved the delivery to 9:30” would need a different check from “9:30 is the latest arrival that makes the schedule fit.” The second is a calculation you can inspect. The first claims that an external booking changed, so you would want to see the updated confirmation before telling the installation team to arrive earlier.
This distinction also helps you decide whether a slower mode earned its time. Did it find an overlap you missed? Did it ask for the fact needed to answer? Did the confirmed action match what you authorized? Those are useful improvements. A longer explanation with the same wrong arrival time is more reading.
OpenAI’s September 16 guide to connecting AI usage with business value makes a related point for organizations: examine the kind of work being done and include review and correction time when evaluating the benefit. You can borrow that question without needing an enterprise dashboard. Judge whether the whole task became easier to finish correctly.
Also this week...
ChatGPT is now available inside Word. OpenAI’s September 17 update adds drafting, summaries, selected-text revisions, and formatting through the Microsoft add-in. It is available across ChatGPT plans, including Free, subject to usage limits. If you already work in Word, that may remove some copying between windows; inspect the changed passage in the document before accepting it.
Anthropic is bringing evaluators inside the lab. Its September 18 partnership announcement says Faculty, Accenture’s specialist AI business, will lead work on model evaluations, safeguards, and alignment with internal access. Anthropic will fund the work directly, and reporting and access standards are still being developed. More access could make scrutiny better informed, but this is an announced arrangement, not a set of completed independent findings.
💡 Suggested how-to: repair one disappointing answer
Here is an exercise you can use with a low-stakes task you understand well. It has not been run as a controlled comparison for this issue. Its purpose is to help you find the right correction before changing several things at once.
Start with an answer you rejected and write down the exact defect. “It was bad” is too broad to guide the next attempt. “It scheduled photography before the display was finished” gives you a check you can repeat.
Compare the answer with your original request. If you never supplied the relevant fact, add it. If the fact was present but the answer ignored its relationship to another fact, keep the information unchanged and try an available deeper thinking setting. If the requirements contradict one another, ask for the conflict and the smallest change that would resolve it.
You can adapt this request:
Check this plan against the requirements below. Treat the confirmed times and durations as fixed. Identify any task that starts before its prerequisite finishes. If the plan is impossible, show the conflict and the earliest possible finish. Keep proposed changes separate from confirmed facts, and ask for missing information instead of inventing it.
Then check the output against the defect you named. For the fictional gallery schedule, the answer must recognize noon as the earliest finish and the 30-minute shortfall. A more elegant table does not compensate for missing that conflict.
Keep the result in proportion. One successful attempt is a reason to try that approach again on similar work, not proof that a setting is universally better. If the same failure remains, stop cycling through increasingly elaborate prompts and inspect the underlying facts or work through that step yourself.
🧰 Tools to try
ChatGPT’s thinking control: Use the plan-specific guide to find the options your account supports. Try a deeper setting on a task with interacting constraints, while keeping the request stable enough to judge the difference. You do not need to upgrade a subscription merely to begin checking your current setup.
A plain calendar or spreadsheet: Put each task’s start, duration, and prerequisite where you can see them together. For a short schedule, that may reveal the conflict faster than another conversation. AI can help construct the plan while the visible rows give you an ordinary way to inspect it.
The next time an answer disappoints you, point to the first thing it got wrong. If the missing piece is a fact, supply it. If the facts were there and the relationship was missed, check the thinking setting before trying again. You will have given the next answer a specific job to do.
P.S. An AI wrote these sentences from this week’s primary sources. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.