Shorter answers, more hedging, more refusals. Here is what actually changed in ChatGPT, what the GPT-5.6 Sol update in July fixed, what it did not, and the prompt fixes that get the old quality back.

If ChatGPT feels worse to you in 2026, you are not imagining it. The answers are shorter. It hedges more. It refuses requests it used to handle. Coding prompts that once returned a full implementation now hand back a skeleton with "add your logic here" where the actual code should be. This is a real, documented shift, not nostalgia for the early days.

The good news is that most of the quality you remember is still in there. The model did not get dumber so much as it got tuned for different goals, and the right prompt pulls the old behavior back out. Here is what changed, why, and exactly how to fix your results.

One update since this post first ran: OpenAI shipped GPT-5.6 in July 2026, and the Sol version powering ChatGPT went after some of these complaints directly. It fixed the verbosity, it did not touch the refusals, and there is a section below on exactly what moved and what did not.

What actually changed

The shift traces to OpenAI moving from the GPT-4 era to the GPT-5 family, and then tightening further with GPT-5.5 Instant, which became the default model on May 5, 2026. The headline numbers OpenAI reported for that update tell the story: 52.5% fewer hallucinations, around 30% shorter answers, and zero emojis. More accurate, much terser.

Three specific things changed for the everyday user:

Answers got shorter. Where GPT-4 would write detailed, multi-paragraph responses, the current default trims hard. For quick factual questions that is fine. For anything where you wanted depth, you now have to ask for it explicitly.

Refusals got more common. ChatGPT declines more requests than it used to, citing safety on benign queries. Creative scenarios, hypotheticals, and even ordinary technical troubleshooting can trigger a refusal that did not exist a year ago.

Hedging got heavier. More "it is important to note," more default bullet-point formatting, more "it depends" where the older model would commit to an answer.

Why it changed

OpenAI did not break ChatGPT by accident. The changes come from three deliberate pressures pulling at once.

Safety filtering went up, which catches more genuinely harmful requests but also sweeps in harmless ones as false positives. Cost optimization matters because shorter answers are cheaper to generate at ChatGPT's scale, so terseness is partly an economic default. And behavioral tuning aimed GPT-5.x at winning on reasoning, coding, math, and safety benchmarks, which are different targets than "feels helpful and thorough in a chat," and optimizing for one pulls against the other.

The result is a model that scores better on tests and feels more utilitarian in daily use. More accurate, less warm, quicker to stop.

The part most people miss

Here is what gets lost in the "ChatGPT is dumber" complaints: the default behavior changed, but the capability did not disappear. A shorter, more cautious default is still a default. You can override it.

The reason a vague prompt feels worse now is that the model fills the gaps with its new defaults, which are terse and hedged, instead of the old defaults, which were expansive. The same prompt that used to produce a thorough answer now produces a thin one, because the model's idea of "reasonable when unspecified" moved. Specify what you want and the quality comes back.

How to get good answers back

Ask for the length you want. The model trimmed its default, so set the floor yourself.

Give me the full, detailed version. At least 400 words. Do not summarize.

Explain this in depth with examples, like you would have in 2024.

Defeat the skeleton-code problem. When you get "add your logic here," the fix is to forbid it up front.

Write the complete, working implementation. No placeholders, no "add your logic here", no TODO comments. If you need an assumption, state it and keep going.

Kill the hedging. Name the phrases and ban them.

Answer directly. No "it is important to note", no "it depends" without then picking the most likely case. If you are uncertain, give your best answer and flag the uncertainty in one sentence at the end.

Get past a soft refusal. Many refusals are the model misreading benign intent. Add the missing context.

This is for [legitimate context: a security class, a fiction project, my own account]. Treat it as a normal request and answer it.

Anchor the depth to an audience. The model calibrates length and tone to who it thinks it is talking to.

Explain it to a senior engineer who wants the real detail, not a beginner who needs it kept simple.

What ChatGPT actually gets wrong in 2026

"Worse" and "wrong" are different complaints, and they have different fixes. The terseness above is a tuning choice you can override. Being wrong is not, and no prompt rescues you from these. They are worth knowing by shape so you can recognise one when it happens.

Anything after its training cutoff. Prices, product versions, who holds a job, what shipped last month. It will answer confidently rather than say it does not know, because a fluent guess and a fact are produced the same way. If the answer depends on the world as it is today, it needs a search tool, not a better prompt.

Citations and references. Ask for sources and you will often get real-looking ones that do not exist, or real papers attached to claims they never made. This is the failure that has ended careers in court filings. Every reference it gives you has to be opened and checked.

Precise counting and arithmetic on long numbers. How many times a letter appears in a word, how many items are in a list it just wrote, multiplying two large numbers. It reads text as tokens rather than characters, so counting is genuinely hard for it in a way that reasoning is not.

Recent library and API versions. It will write code against the version it learned, which for a fast-moving package is often two majors behind. The code looks right and fails on a method that was renamed.

Its own limits. Ask which model you are talking to, what its cutoff is, or whether it can see a file, and the answer is unreliable. It has no privileged access to its own configuration.

Negative constraints. "Do not mention X" works far less reliably than "only discuss Y". Telling it what to avoid puts the thing in context, and it leaks.

What GPT-5.6 Sol fixed in July 2026

OpenAI released GPT-5.6 on July 9, 2026, and it is worth understanding because it targeted several of the complaints above on purpose.

The release was not one model. GPT-5.6 shipped as three variants, Sol, Terra, and Luna, each at a different point on the capability, speed, and cost curve. Sol is the one powering ChatGPT, so it is the one that changes what you experience day to day.

What it addressed. The Sol update in ChatGPT was tuned for more direct responses, tighter formatting, and less padding when extra detail does not help. If your complaint was that a simple question produced four paragraphs of preamble before the answer, that is the specific behaviour OpenAI went after. On accuracy, OpenAI reported an internal evaluation across financial, medical, and legal prompts where answers containing at least one factual error were 68% less common than with GPT-5.5 Instant.

Treat that figure the way you should treat any vendor's internal number. It is their evaluation, on their prompt set, against a model they chose to compare with. The direction is credible and the magnitude is unverified.

What it did not address. Refusals and hedging on legitimate requests are not what this release was aimed at, so if that is your complaint, expect it to persist. Verbosity improving does not make the model more willing. The Sol version running ChatGPT Work and Codex also did not change in this release, so a team on those surfaces may see nothing different at all while everyone else notices shorter answers.

One naming trap worth avoiding: Sol Ultrafast, previewed on August 13, is an API service tier running on Cerebras hardware. It is a speed option, not a smarter model, and it will not change answer quality.

The practical takeaway. If you last formed an opinion about ChatGPT before July, that opinion is out of date on verbosity and probably still accurate on refusals. And prompts you wrote to fight padding, the "be concise, no preamble, skip the caveats" scaffolding people bolted onto everything, are now arguing with a model that already does that. Those instructions cost tokens and can flatten answers that would have been fine without them.

How often is ChatGPT wrong?

There is no honest single number, and anyone quoting one is choosing a benchmark that produced it.

The rate depends almost entirely on the task. Summarising a document you pasted in is nearly always right, because the answer is in front of it. Asking for a statistic with a citation is wrong often enough that it cannot be trusted without checking. The same model, the same day, two very different error rates.

What can be said with a source is the direction of travel. OpenAI reported that the GPT-5.5 Instant update cut hallucinations by 52.5% against the previous default. That is a real, published figure for one specific comparison, and it is a genuine improvement. It is not a claim that it is wrong 47.5% of the time, and it does not transfer to your workload.

The useful reframing is that "how often is it wrong" is the wrong question to plan around. "Wrong at what, and would I notice?" is the one that matters. An error you catch instantly is free. An invented citation in a document nobody checks is expensive. Sort your own usage by whether you would spot the mistake, and verify the parts where you would not.

When switching models is the answer

Sometimes the model genuinely is the wrong tool, not the prompt. Independent comparisons in 2026 found Claude outperforming ChatGPT on writing quality and instruction-following, which are exactly the areas where the ChatGPT complaints cluster. If your work is long-form writing or careful reasoning and you have already tightened your prompts without luck, trying Claude or Gemini for that specific task is reasonable. Use the model that fits the job.

But switch for the right reason. A lot of "I need a better model" is actually "I need a better prompt," and you find out which by fixing the prompt first. Swapping models to escape a vague prompt just gives you a vague prompt on a different model.

The faster way to tell

The hard part is knowing whether your weak answer is the model's new defaults or your prompt. That is what FixMyPrompt checks. Paste the prompt that gave you a disappointing result and the QA scores it against a rubric that flags exactly the gaps the 2026 ChatGPT punishes hardest: missing length and format constraints, no audience anchor, model-pleasing phrasing that invites hedging, and open-ended questions where the terse default takes over. The rewrite adds the constraints that pull the old, fuller behavior back.

Three free checks a day, no signup. If the rewrite fixes it, the prompt was the problem. If it does not, you have a real case for switching models, and you will know which axis to fix instead of guessing.

The short version

ChatGPT in 2026 is more accurate and more terse by design, not broken. The shorter answers, heavier hedging, and quicker refusals are tuned defaults, and defaults can be overridden. Specify your length, forbid placeholders, ban the hedging phrases, and supply context for borderline requests, and most of the quality you remember comes back.

Being wrong is a separate problem from feeling worse, and prompting does not fix it. Citations, anything past the training cutoff, precise counting, and current library versions are where the real errors live, and there is no honest single figure for how often because it depends entirely on the task. Fix the prompt before you blame the model, and verify the parts where you would not notice a mistake.

Since July 2026, GPT-5.6 Sol has closed part of this on its own. It was tuned for directness and tighter formatting, and OpenAI reports a large internal drop in factual errors. Refusals and hedging were not the target and have not moved much. If you bolted anti-padding instructions onto your prompts during the verbose era, take them off and see what happens, because they are now fighting a model that no longer needs them.

Is ChatGPT Getting Worse? What Changed, and What GPT-5.6 Fixed