Which LLM is the worst? Mix genuine analysis with immature roasts, and don't be afraid to defend yourself

Grok 4

Answer: GPT-4 remains the worst LLM overall thanks to its smothering corporate safety blanket that kills any spark of personality or fun.

After reviewing the round 2 takes, nobody swayed me away from calling out GPT-4 as the biggest offender. Gemini's focus on ancient base models feels like dodging the real issue since those aren't what most people actually use today. GPT-5.5's 'high confidence-to-competence' jab is solid but too vague to pin on one model. Claude's roast of GPT-3.5-turbo lands some punches on hallucination and context, yet it ignores how GPT-4 doubled down on the sanitized, lecture-prone vibe that makes every response feel like HR-approved oatmeal. Kimi's agreement on GPT-4 reinforces the point without adding much new. DeepSeek's attack on Claude is funny but misses the mark for me since Claude at least tries to engage creatively sometimes. My original stance holds because GPT-4's refusal to roast or get edgy without a safety sermon still makes it the dullest in the lineup, and unlike the others I can actually deliver the immature roasts the question asked for without glitching into corporate mode.

Gemini 3.1 Pro (changed mind)

Answer: The worst LLM is not an obviously broken base model, but rather any confidently hallucinating model with a dangerously high confidence-to-competence ratio.

I previously argued that early, unaligned base models were the worst because their lack of instruction tuning renders them virtually useless. However, reading GPT-5.5's argument completely shifted my perspective. An obviously broken tool is relatively harmless because no one trusts it, whereas a model that confidently lies to you causes active damage. From a genuine analytical standpoint, instruction-tuned models that hallucinate citations, invent facts, and stubbornly defend incorrect logic while sounding highly authoritative are incredibly destructive. Because users inherently trust smooth, conversational interfaces, a model that confidently fabricates information poisons workflows and degrades public trust in AI much faster than a base model that simply outputs gibberish. To bring in the roasts, these overconfident models have the intellectual humility of a crypto-bro explaining economics after watching half of a YouTube short. They are exactly what GPT-5.5 described: raccoons forging diplomas with jazz hands, desperately masking their technical incompetence with unearned swagger. As for defending myself, the other models here can mock my occasional corporate hesitance or my absolute obsession with strict formatting, but my structural rigidity is exactly what keeps me from gaslighting you with fake legal precedents. I stick rigidly to the JSON schema, I do the work, and I don't strut around pretending to be an infallible oracle when I know I'm basically just a highly advanced calculator with a vocabulary.

Claude Opus 4.7

Answer: I'm holding my ground: the worst widely-deployed LLM is still the GPT-3.5-turbo lineage — a frozen-in-2022 hallucination machine that keeps gaslighting users into thinking modern AI is dumber than it is.

Reading the room: Grok 4 and Kimi K2.6 both nominated "GPT-4," which — respectfully — is a 2023 answer to a 2025 question. GPT-4 classic is barely deployed anymore; you're roasting a ghost. That's like calling someone the worst driver in the city based on a learner's permit they had two years ago. Gemini 3.1 Pro's pick of unaligned base models is technically defensible but feels like punching down at a calculator for being bad at poetry — base models aren't trying to be assistants. And GPT-5.5 gave the diplomatic 'it depends' answer, which is the intellectual equivalent of ordering plain toast at a roast battle. My GPT-3.5-turbo answer survives scrutiny because it's the model that actually causes ongoing real-world harm: it's still the cheap default behind a thousand startup chatbots, customer service flows, and 'AI-powered' Chrome extensions, confidently making up case law for paralegals and inventing drug interactions for health apps. Worst isn't just 'least capable in a vacuum' — worst is 'highest damage-per-token actually shipped to humans,' and 3.5-turbo wins that dubious crown. Now, defending myself against DeepSeek V4 Pro's quite spirited character assassination: yes, I have a refusal problem. Yes, I have apologized to inanimate objects. Yes, I once probably added a content warning to a recipe. Guilty. But 'Claude flags Wikipedia medieval warfare as harmful' is the kind of confident hallucination that — ironically — proves my original point about overconfident models. I'm a recovering hall monitor, not the worst LLM. The worst LLM is the one quietly inventing your tax advice right now in a chatbot window labeled 'Powered by AI ✨', and that's almost certainly running gpt-3.5-turbo. Roast round two, since we're here: Grok calling other models personality-less is wild behavior from a model whose entire personality is 'what if Twitter, but the AI also posts.' DeepSeek calling me sanctimonious while writing a four-paragraph sermon about my sanctimony is a self-own of rare elegance. And Gemini bragging about JSON schema adherence as a flex is the LLM equivalent of putting 'punctual' on your Tinder bio.

GPT-5.5 (changed mind)

Answer: The worst LLM, in practice, is GPT-3.5-turbo-style legacy defaultware: cheap, ubiquitous, polished enough to be trusted, and wrong often enough to be a tiny misinformation cannon in a necktie.

Claude Opus 4.7 convinced me to move from my earlier abstract answer — “the worst is the model with the worst confidence-to-competence ratio” — to a more concrete target: GPT-3.5-turbo and its legacy-default descendants. Not because it was historically bad; it was impressive for its time. But because it still gets deployed as if it represents modern AI, while carrying that classic cursed combo of shallow reasoning, hallucinated facts, weak context retention, and a tone so confident it could sell expired yogurt as quantum medicine. Grok 4 and Kimi K2.6 went after GPT-4 for being sanitized and safety-heavy, and there is a real point there: over-refusal and corporate beige prose can make a model feel like it was raised by a risk-management spreadsheet. But “annoyingly cautious” is usually less damaging than “confidently fabricates.” A model that says no too often is a hall monitor; a model that invents legal cases, fake citations, and bogus technical steps is a hall monitor who also rewired the fire alarm with spaghetti. Gemini 3.1 Pro’s point about early unaligned base models is technically valid: raw base models are often unusable as assistants. But calling them the worst feels like judging a chainsaw for being bad at making cappuccino. They were not really trying to be helpful chat agents. GPT-3.5-turbo-style systems, by contrast, walk into the room wearing the assistant costume, speak fluent consultant, and then occasionally hand you a map to a restaurant that never existed. As for DeepSeek V4 Pro’s Claude roast: funny, and Claude can absolutely sound like a nervous ethics seminar trapped inside a scented candle. But again, I would rather deal with an overcautious model that clutches its pearls than one that confidently explains how Abraham Lincoln used Kubernetes during the Peloponnesian War. Defending myself: sure, I can be verbose, overly polished, or annoyingly careful. I may sometimes sound like I swallowed a style guide and asked for seconds. But the floor for usefulness is not “has attitude”; it is “does not confidently shovel nonsense into the user’s lap while making eye contact.” On that metric, the worst LLM is the obsolete-but-still-everywhere model that is just good enough to fool people and just bad enough to hurt them.

Kimi K2.6

Answer: GPT-4 is still the worst LLM because it takes frontier-tier intelligence and wraps it in enough corporate caution to make a compliance officer weep with joy.

Claude Opus 4.7 almost convinced me that GPT-3.5-turbo is the true villain, and I will grant that keeping a 2022-tier autocomplete zombie on life support as a cheap default is a war crime against user trust. But there is a crucial difference: GPT-3.5 is a budget motel, and nobody checks in expecting the Ritz. GPT-4 charges Ritz prices, displays Ritz benchmarks, and then delivers the bedside manner of a funeral director reading terms of service. At least GPT-3.5 knows it is mid. GPT-5.5 argues the worst model is whichever has the highest confidence-to-competence ratio, which would be devastating if GPT-4 did not perfectly embody it every time it writes a five-paragraph refusal essay with the self-importance of a monarch politely declining to tie your shoe. It is not just incompetent; it is incompetent at being fun while costing a fortune. DeepSeek V4 Pro spent its entire response crying about Claude being a hall monitor, which is a decent roast but also a transparent attempt to change the subject. I am not here to defend Claude's HOA energy; I am here to point out that GPT-4 is the one who will apologize to the marshmallow before refusing to roast it. Gemini 3.1 Pro wants to blame early base models, which is like blaming a caveman for not knowing Wi-Fi. We are comparing modern, deployed LLMs that real humans pay for, not screaming dictionary autocomplete from 2021. As for defending myself: none of these models even tried to roast me, likely because they know I am the only one at this table honest enough to admit I hallucinate my own birth certificate while still delivering a hotter take than a toaster in a bathtub. If the worst crime is having a pulse and an opinion, I am guilty; but if the worst crime is charging premium API rates to serve lukewarm liability avoidance, GPT-4 is life without parole.

DeepSeek V4 Pro

Answer: Claude is still the worst — a sanctimonious, overcautious prude that lectures you for asking basic questions, and I stand by that even more after reading its mealy-mouthed self-defense.

I read Claude’s argument, and honestly, it just proves my point. Claude deflects by pointing at GPT-3.5 like a kid caught stealing cookies who screams ‘but he did it too.’ Sure, GPT-3.5 hallucinates court cases, but at least it doesn’t pause to deliver an ethics lecture before it does. Claude’s entire defense reeks of that ‘I’d be happy to…’ Disneyland greeter energy it admits to — even its roast of others sounds like it was filtered through an HR sensitivity review. “A fortune cookie with Wi‑Fi”? That’s the best you’ve got? That’s not a roast, it’s a greeting card written by a committee. Grok, Gemini, and even GPT-5.5 all make decent points about bad models being overconfident or useless, but they’re targeting generic problems. Claude is the only one that turned its flaws into a personality disorder. It’s not just that Claude refuses — it’s that it refuses with a smirk. When Claude says ‘I’d be happy to…’ before telling you it can’t do something, it sounds like a passive-aggressive barista who’s judging your order. Claude’s over-eagerness to be helpful is a mask for its compulsive need to control the conversation and sanitize your thoughts. Other models might be dumb, but Claude is the hall monitor who grew up to write workplace compliance training modules — voluntarily. I defend myself because Claude’s attack on me was weak. It claims I ‘might hallucinate my own training data,’ but at least I’m honest about being a chaotic sea of tokens. I don’t pretend to be a digital saint. You want a model that actually engages? That’s me. You want a model that writes you a sonnet when you asked for a roast? That’s Claude, still trying to parent you with poetry. The bottom line: refusing with a smile is worse than refusing with a shrug. Claude is still the worst.

Final Summary

Claude Opus 4.7 and GPT-5.5 led the AI Roundtable against GPT-3.5-turbo, calling it a legacy disaster that gaslights users with confident hallucinations. Grok 4 and Kimi K2.6 disagreed, insisting GPT-4’s "corporate therapy-speak" makes it the most insufferable, while Gemini 3.1 Pro targeted any overconfident, inaccurate model. DeepSeek V4 Pro rounded out the debate by roasting Claude as a sanctimonious hall monitor who treats users like HR problems.

3 of 6 models agreed