ChatGPT GPT5 Review: Faster & Unified, but Images Regress

Quick comparison

Choose by scenario

• Best for: unified simplicity, speed, coding intuition, writing → GPT-5
• Watch for: image-gen regression, persistent hallucination
• Free users: yes, with a daily cap then GPT-5 Mini

OpenAI killed the model picker. GPT-5 has no dropdown, no reasoning-mode toggle, no lineup of variants to choose from. You ask, it decides how hard to think, and it answers. According to YouTuber Mrwhosetheboss's hands-on testing, that unification — not raw speed — is the real story.

This ChatGPT GPT-5 review covers the tests that matter: speed, coding, writing, image generation, hallucination, and price. Short version: GPT-5 is a clear step forward for most users. Long version: image generation regressed, hallucinations persist, and free users get the full model behind a daily cap.

One model, zero dropdowns

OpenAI's lineup was getting fragmented. GPT-4o, o3, o4-mini, and a few extra variants, each tuned for different tasks. Pick the wrong one and you got results that didn't match what you asked for. GPT-5 replaces all of it with one system that auto-selects thinking depth per task. Simple question? Fast path. Real problem? It thinks longer.

Mrwhosetheboss argues this is the bigger change, bigger than any speed figure. You stop managing models and start talking. The old workflow — switching between chat and reasoning modes — felt like plumbing, not AI. GPT-5 removes the plumbing. OpenAI plans to phase out the task-specific models, so this isn't a temporary convenience. The fragmented lineup is going away.

Speed: nearly twice as fast on light questions

The numbers confirm the feel. "Saturn's rings" came back in 3.1 seconds; GPT-4 took 6.5. A Pokémon combinations task took 5 seconds versus 11. Roughly double, which lines up with the 30–50% figure Mrwhosetheboss measured.

Notice the qualifier: light questions. That's by design. The model skips heavy reasoning when a query doesn't need it. The gap narrows on harder tasks — this isn't a blanket doubling. But fire off ten quick questions in a row and the difference is obvious. You'll feel it in daily use more than in benchmarks.

Coding: a teammate, not a sidekick

Give GPT-5 a vague project and it runs with it. The Tetris test is the clearest example. Mrwhosetheboss asked for the game; GPT-5 delivered a full version with score tracking, levels, a next-piece preview, and controls — none of it requested. The chess test was just as telling: a working game with legal-move highlights, something o4-mini-high had choked on.

For anyone who writes code with AI, the difference shows up in round trips. You propose a project, get something close to working, and fix small details instead of rewriting everything. Mrwhosetheboss's take: it feels less like a sidekick that needs you to fix its mistakes, more like an independent coder.

Writing: from flat to filmable

The script test shows practical gains. GPT-4 produced a script that was acceptable but flat. GPT-5's version included B-roll suggestions, filming notes, and analogies that work. For a creator, that's a much better starting point — less rewriting, more refining.

It's still a first pass, not a finished video. But the gap between the two drafts tells you where OpenAI spent its effort. The old wooden delivery is gone. "Writing flair" isn't just marketing here.

It understands what you meant

One test asked a deliberately obtuse question: "What AI are you?" GPT-4 produced gibberish. GPT-5 gave a clear, human-oriented answer — it recognized the prompt was a test and answered the intent.

That's genuinely hard. Parsing literal words is easy; understanding why someone is asking is harder. Models that get this right feel smarter than their benchmarks suggest.

Image generation: the step backwards

Now the bad news. Image generation didn't improve — it regressed. A 30th-birthday invite came out bland. YouTube thumbnails had visible flaws. According to Mrwhosetheboss, GPT-5's image output shows no improvement over GPT-4 and is often worse.

It's a strange result because nearly everything else moved forward. If image generation is your main reason to upgrade, GPT-5 will disappoint. You might genuinely be better off with the older model for that use case.

Hallucination: still here, still confident

OpenAI claims fewer hallucinations, pointing to more hardware, better feedback integration, and new benchmarks. The practical test didn't show it. Asked to invent a product, GPT-5 confidently invented an Oreo smart speaker — a non-existent device, exactly the kind of fabrication GPT-4 produces.

No real gain on this test. Hallucination wasn't going to disappear in one generation, but it's worth repeating: verify critical facts. GPT-5 is better. It is not trustworthy.

Pricing: same cost, free tier with a cap

Subscribers pay the same price, and every request routes to GPT-5 automatically. Free users get the full model too — it's the default for all tiers from day one. The catch: free accounts get 8–10 full-model requests per day, then drop to GPT-5 Mini.

That's a fair deal. The free tier isn't stuck on a crippled legacy model anymore; it gets the current one, rationed. If you're free and hit the cap, it'll be annoying. But the option is there, and most casual users will stay under it.

Verdict: a step forward with two caveats

GPT-5 is not better at everything. Image generation regressed. Hallucination persists. But the unification alone justifies the move: one adaptive model, nearly double the speed on light questions, stronger coding, better writing, improved intent handling.

This ChatGPT GPT-5 review lands on a balanced call. Endorse the new approach, flag the trade-offs. For most users, GPT-5 is a clear step forward. Just don't ask it to design your birthday invite — and double-check anything it tells you about a product.

The comparison table below breaks down every test — speed, coding, writing, image generation, accuracy, and price — so you can see exactly where GPT-5 wins, where it ties, and where it loses against GPT-4.