Why Are My ChatGPT Answers So Short? A 2026 Diagnostic

TL;DR: ChatGPT answers are short because something lowered the model's default level of detail: a personality or verbosity setting, a saved instruction you forgot about, a low reasoning level, or a prompt that implies a small deliverable. Word counts do not fix it. Naming the content you require does.

Why are my ChatGPT answers so short?

Because the length of a ChatGPT answer is decided by four things, and only one of them is the sentence you typed. The other three are a default detail level baked into the model and your settings, any standing instruction stored on your account, and how much of the response budget went to invisible reasoning before the visible text began.

That matters because the five causes look identical from the outside. Every one of them produces the same symptom: a thin answer where you expected a thorough one. They need completely different fixes, so guessing costs you more time than checking.

Work through them in this order.

The fix for most of them is the same, and it is the opposite of what most advice tells you to do. Skip to the fix if you already know which cause is yours.

Cause 1: is there a ChatGPT setting that makes answers short?

Yes, several. None of them is labelled "length".

The first is Personality, in Settings then Personalization. OpenAI's help article documents seven presets: Default, Cynical, Efficient, Friendly, Professional, Quirky and Candid. The gap between them is not subtle, and OpenAI's own side-by-side examples make the point better than any argument. Given the prompt "How should I spend a rainy day indoors?", the Efficient personality answers: "Read, cook, clean, repair, or work on indoor hobbies. Organize files, learn online, or exercise at home." Given a coding request, the same preset is documented as producing no intro at all. If you set Efficient months ago and forgot, you have been reading a terse model ever since.

The second is Characteristics, in the same panel. These are plus and minus nudges applied across chats. OpenAI's help article says they let you adjust "brevity, tone, formatting, emoji level, etc.", and that is worth reading closely, because the four characteristics the article actually documents with examples are Warm, Enthusiastic, Headers & Lists and Emojis. Brevity is named in the introduction; no brevity control appears in the documented list. The feature is also described as rolling out gradually, so it may not be in your account yet.

Headers & Lists deserves its own note. Turning markdown down makes ChatGPT "rely more on paragraphs", in OpenAI's phrasing. That barely changes the amount of information, but it changes how much screen the answer occupies. A fair number of "ChatGPT got shorter" complaints are really "ChatGPT stopped bulleting", which is a formatting change, not a depth change.

The third is the model default, and this one moves under you. OpenAI's own migration guidance for developers states plainly that GPT-5.6 tends to be more concise by default than GPT-5.5, and advises checking whether standing instructions like "Be concise" are still helping, because they "can sometimes make responses too brief". You did not change anything. The default did.

Cause 2: does your prompt imply a short answer?

Very often, yes, and this is the cause people are least willing to consider.

A model reads your prompt for the deliverable it names, not for the effort you hoped for. "What is a retention cohort?" names a definition. Definitions are short. "How do I price a retainer?" names an answer, and answers are short. Neither prompt asked for a document, so neither got one. The word "quickly", the phrase "in short", and the verbs "summarize" and "explain" all set a register, and the model honours the register you set.

The quieter version of this is a prompt with nothing to be long about. "Write about employee onboarding" contains no audience, no company size, no situation and no constraint. There is no material in it. The model has to invent every specific it uses, and a model that is unsure what you want tends to hedge toward the safe middle, which is short and generic. Adding context is not politeness. It is raw material. Our guide on writing better ChatGPT prompts works through the components that make a prompt answerable in the first place.

There is also a cost to contradicting yourself. OpenAI's GPT-5 prompting guide notes that "poorly-constructed prompts containing contradictory or vague instructions can be more damaging to GPT-5 than to other models, as it expends reasoning tokens searching for a way to reconcile the contradictions rather than picking one instruction at random." A standing custom instruction that says "keep answers brief" plus a prompt that says "go into as much depth as possible" is exactly that kind of conflict. Something has to lose, and you will not be told which.

Cause 3: is reasoning eating the answer before it starts?

This is the cause almost nobody checks, and in the API it is the only one that can delete the answer entirely.

Reasoning models generate tokens to think with before they generate tokens for you. Those reasoning tokens are invisible, they occupy space in the context window, and OpenAI bills them as output tokens. The max_output_tokens parameter caps everything the model generates, reasoning included, plus non-visible formatting tokens that never appear in the message content.

The failure mode follows directly. In OpenAI's words: if the generated tokens hit the limit "you'll receive a response with a status of incomplete... This might occur before any visible output tokens are produced, meaning you could incur costs for input and reasoning tokens without receiving a visible response." A short answer, in this scenario, is a truncated one, and a truly unlucky one is empty. OpenAI's token-counting guide adds the practical rule: "Leave headroom in these limits when you need a specific amount of visible output."

In the ChatGPT app you do not set token limits, but you do set reasoning level, and as of August 2026 that control is a slider. OpenAI documents Instant, Medium, High, Extra High and Pro on eligible paid plans, powered by GPT-5.6 Sol, with GPT-5.6 Sol Pro behind Pro. Free and Go accounts get GPT-5.6 Luna and a Think option for harder questions. There is also an automatic mode: Settings, then General, then "Higher intelligence", which lets ChatGPT reason more on complex requests while Instant stays selected.

Cause 4: did you tell ChatGPT to be brief and then forget?

Custom Instructions are the single most common cause of chronic brevity, precisely because they are invisible at the moment you notice the problem. They live in Settings, then Personalization on web and desktop, or Settings, then Customize ChatGPT on iOS and Android, behind an "Enable customization" toggle. OpenAI documents the limits as 1,500 characters on Free and Go, and 5,000 on Plus, Pro, Enterprise, Business and Education. A "no fluff, keep it tight" line written a year ago is still running today.

Memory is the sharper edge. OpenAI's memory FAQ states that saved memories "are part of the context ChatGPT uses to generate a response" and that "unless you delete them, saved memories are always considered in future responses." It also says ChatGPT "may save those details as a memory without you needing to ask." So a passing "that was too long, keep it shorter" in one conversation can quietly become a standing preference across all of them. The newer memory system compounds this: OpenAI notes that memory is "a continually updated synthesis of context from your past chats, which may be broader than what can be shown as individual items in a summary". Scanning the memory list is therefore not proof that nothing is stored.

Both of these behave like an always-on system prompt that you cannot see in the chat window. If you want the mental model for how that layer overrides what you type in the message box, system prompts versus user prompts covers the hierarchy.

The test takes a minute. Temporary Chat does not use saved memories, which makes it a clean control.

Run this identical prompt twice: once in a normal chat, once in a Temporary Chat.

"Explain how compound interest works to someone who has never
invested. Include one worked numeric example over 10 years,
one common misunderstanding, and one thing that changes the
answer if the deposits are monthly rather than annual."

Longer in Temporary Chat  -> the cause is stored (memory or custom instructions)
Same length in both       -> the cause is your prompt or your reasoning level

Cause 5: why doesn't "write 2,000 words" work?

Because the model is not counting.

Text generation is one token at a time, and a token is not a word. There is no running total the model consults as it writes, no internal odometer ticking toward 2,000. The evidence for this is in OpenAI's own tooling: counting tokens requires either a separate local library or a dedicated API endpoint, and that endpoint counts input tokens, before the call. Output counts arrive after generation, in the usage object. Nothing in the loop gives the model a live word count.

So a number in your prompt is a stylistic signal, not a constraint. The model has learned what a request for "2,000 words" tends to look like and aims in the general direction, which is why you get 700 when you asked for 2,000, and 2,600 when you asked for 500. That failure is worth its own explanation, and we wrote one: why AI ignores your word count.

The useful conclusion is not "you cannot control length". It is that you have been aiming at the wrong variable.

The fix: depth instructions beat length demands

Here is the counter-intuitive part. You get length by requiring content, not by requesting words.

A word count asks the model to hit a number it cannot measure. A depth instruction hands it a list of things that must exist in the finished answer. The second is checkable. The model can tell whether it has covered point four; it cannot tell whether it has written 1,900 words or 2,100. Give it a completion test it can actually pass and length arrives as a by-product.

Compare the two prompts directly.

❌ Length demand

Write 2,000 words about employee onboarding.

✅ Depth instruction

Write an onboarding guide for a 12-person marketing agency hiring
its first full-time project manager.

Cover each of these five points, with a worked example drawn from
that specific setup:

1. What must exist before day one: accounts, documents, first-week calendar
2. The first 48 hours, hour by hour
3. Who owns what in weeks 1 to 4, named by role
4. The three things a new agency PM most often gets wrong
5. How the manager checks at day 30 whether it worked

For every point, include: one paragraph of explanation, one worked
example using this agency, and one failure mode.

End with a checklist the manager can print. No preamble, no summary.

Count the second prompt again. Five points, three required elements each, plus a checklist: sixteen things that must appear. There is no way to satisfy that in four paragraphs. Length stopped being the request and became the arithmetic.

This is also what OpenAI recommends. Its GPT-5.6 prompting guidance tells developers to identify "the information the model must preserve and the detail it can omit", and gives this as a model instruction: "Keep all required facts, decisions, caveats, and next steps. Trim introductions, repetition, generic reassurance, and optional background first." That is a content contract, not a size request, and it works in both directions. The same technique that produces long answers produces reliably short ones.

A template you can paste

Swap the bracketed parts. The structure is what does the work.

[Task, with the specific audience and situation]

Cover each of these points:
- [point 1]
- [point 2]
- [point 3]
- [point 4]

For each point, include:
- a direct answer in one or two sentences
- one worked example using [my specific context]
- one thing that commonly goes wrong, and how to tell

Then: [a closing artifact, e.g. a checklist, a table, a decision tree]

Do not write an introduction. Do not summarise at the end.

The two "do not" lines matter more than they look. Preamble and summary are the parts a model reaches for when it is padding, which means they are also the first things it produces when it has nothing specific to say. Banning them forces the substance to carry the answer.

The length instructions that actually do work

Not every length instruction is useless. The ones that work describe structure, because structure is countable and word totals are not.

✅ Works: countable structure

Answer in exactly three paragraphs.
Paragraph 1: the recommendation.
Paragraph 2: the two strongest reasons for it.
Paragraph 3: the strongest objection, and your response to it.

✅ Works: bounded units

Give me 8 options. One line each, maximum 15 words.
No explanation, no ranking.

❌ Unreliable: uncountable targets

Around 800 words.
Be comprehensive.
Go into as much depth as you can.

"Three paragraphs" is verifiable while writing. "Eight options at fifteen words each" is verifiable per item, which is why per-item limits hold up far better than whole-document ones. "Around 800 words" is verifiable only after the fact, by you.

Choosing the shape of the output before you ask for it is a skill in its own right, and it changes the answer more than tone or phrasing does. How to decide what output format you actually need walks through picking between prose, table, list and structured output for a given job.

A 60-second diagnostic

Run these in order and stop at the first one that changes the answer.

1. Open Settings, then Personalization. Check your Personality. If it is Efficient, that alone explains chronic terseness. Read your Custom Instructions in full, out loud if necessary, looking for anything that means "be brief".
2. Open your memory list and look for saved preferences about length or tone. Remember that the list may not show everything the synthesis contains.
3. Run the control prompt in a Temporary Chat. Longer means the cause is stored. Same length means it is the prompt or the reasoning level.
4. Check your reasoning level. If you were on Instant, try Medium or High and resend the same prompt unchanged.
5. Reread your prompt for the deliverable it names. Does it ask for a definition, or a document? Does it contain the words "quick", "brief", "summarize" or "in short"?
6. Only now, rewrite. Replace any word count with a list of required points and a required element count per point.

If you get through all six and the answers are still thin, the problem is usually upstream of length. Our post on why your ChatGPT answers are bad covers the failures that a longer answer only makes worse.

When a short answer is the correct answer

Worth conceding: sometimes brevity is not a bug you should fix.

If your question has one right answer, a long response is padding, and padding is where hallucinated detail gets in. A model asked to fill 2,000 words on a topic that warrants 200 will find something to say, and the something it finds is the least reliable part of the output. Length is not a proxy for quality, and forcing it can lower quality outright.

The honest version of this article's advice is narrower than "make ChatGPT write more". It is: decide what has to be in the answer, require exactly that, and let the length be whatever that requires. Some days that is 1,800 words. Some days it is a sentence, and the sentence is right.