GPT-5 Review: We Tested It for 30 Days — Here's the Real Verdict

We've been using GPT-5 since early access opened up, and the short version is: it's genuinely a step change, not just an incremental update. The longer version is more nuanced. GPT-5 is significantly better at reasoning and coding, noticeably better at following complex instructions, and occasionally makes the same kinds of errors GPT-4 made. Here's everything we found after 30 days of real-world testing.

What's Actually New in GPT-5

The headline number OpenAI touts is PhD-level performance on standardized benchmarks. On the MMLU benchmark (a 57-subject knowledge test), GPT-5 scores 92.3% compared to GPT-4's 86.4%. On MATH (a college-level math test), it jumps from 67% to 88%. Those are real, meaningful improvements — especially for anyone using AI for technical work.

Less publicized but arguably more important: GPT-5's instruction following is dramatically better. If you've ever had GPT-4 ignore a specific formatting requirement halfway through a long response, you'll notice GPT-5 is much more reliable. In our testing, it maintained correct JSON formatting across outputs 40+ lines long — something GPT-4 would occasionally corrupt.

Coding Performance: A Real Upgrade

This is where GPT-5 shines most clearly. We threw 50 real-world coding problems at it — bugs from production codebases, system design questions, performance optimization tasks. GPT-5 solved 44 of them correctly on the first attempt. GPT-4 Turbo managed 31 of the same problems.

More importantly, GPT-5's explanations are better. When it fixes a bug, it explains not just what was wrong but why the original code would fail under edge cases. That context is what separates a useful coding assistant from a one that just pattern-matches.

For multi-file refactoring tasks, GPT-5 maintains consistency across files significantly better. It's not perfect — we still saw it introduce naming inconsistencies on very large tasks — but it's a meaningful improvement.

Writing and Creative Work: Good, Not Revolutionary

For writing assistance, GPT-5 is better but the gap is smaller. Creative writing feels slightly more natural with less repetitive phrasing. Long-form content like reports and essays is more coherent and better structured. But anyone using GPT-4 for writing was probably already getting solid results.

One area where GPT-5 stands out: research-style tasks where you need to synthesize multiple perspectives. Ask it to write a balanced analysis of a complex topic and it's genuinely more thoughtful about representing different viewpoints accurately.

Where GPT-5 Still Struggles

Hallucinations haven't disappeared. GPT-5 is less likely to confidently state wrong facts than its predecessor, but it still does it. In our testing, it hallucinated details about obscure software libraries and got dates wrong on recent events. Never use GPT-5 for anything where factual accuracy is critical without verification.

The context window is 128K tokens — same as GPT-4 Turbo. There's no 1M+ context option like Gemini offers, which matters for very large document processing. For most use cases this is fine, but if you're processing entire codebases or research libraries, Gemini still wins on raw context capacity.

Is It Worth the Upgrade Cost?

GPT-5 is available via ChatGPT Plus ($20/month) for standard access, with higher usage limits on the Pro tier ($200/month). If you're currently on the free tier, the jump to Plus is absolutely worth it — the difference between GPT-3.5 and GPT-5 is enormous. If you're already on ChatGPT Plus, upgrading to the full GPT-5 access (which may require the Pro plan) depends on your use case.

For developers and technical users: yes, upgrade. The coding improvements alone justify the cost if AI is part of your daily workflow. For general writing and productivity: the free tier with GPT-5 access (rolling out gradually) is probably sufficient.

Conclusion

GPT-5 is OpenAI's best model by a clear margin, and it's one of the best AI models you can use today. The coding and reasoning improvements are real and meaningful. The writing improvements are incremental but genuine. The hallucination problem is reduced but not solved. For anyone serious about getting the most out of AI tools, GPT-5 is the current benchmark — but keep Gemini and Claude in your toolkit for their specific strengths.

Frequently Asked Questions