GPT-5 is now available. After the beta at the end of 2025 and the staggered rollout, OpenAI has opened the model to all Plus plans and to the API in Tier 1+ plans. The question that matters: what changes for your daily use and is it worth moving from GPT-4?
We’ve spent two weeks using GPT-5 on real tasks from the editorial team — writing, data analysis, code, agents — and this is the summary without marketing.
What is GPT-5 (and what it’s not)
GPT-5 is not a single model: it’s a unified family with three variants that the system automatically chooses based on the task:
- GPT-5 Instant — quick responses for general conversation, replaces GPT-4 as the default model.
- GPT-5 Thinking — deep reasoning, under the hood it uses the equivalent of “o3” from last year.
- GPT-5 Pro — extended version for Pro subscribers (200 €/month) with greater context and depth.
For the first time, you don’t have to choose the model manually. The system does it based on the prompt — and the era of “should I use o3 or 4?” is over.
Real improvements we’ve noticed
Structured reasoning
Tasks with several steps (analyze these data, propose 3 hypotheses, check each one with this dataset and give me the winner) are completed without having to divide them into sub-prompts. GPT-4 used to lose track after 3-4 steps; GPT-5 keeps the chain.
Lower hallucinations, but not zero
According to OpenAI, hallucinations decrease by ~45% vs GPT-4 in factual questions. In practice, this is noticeable especially in code (it generates fewer non-existent imports, fewer invented functions) and in responses with numerical data. It still invents if the question is ambiguous: if you ask for data it doesn’t know, it assumes and fills in.
Significantly better programming
In SWE-bench Verified (standard benchmark for coding) GPT-5 scores 74.9% vs 30.8% of GPT-4. This is noticeable in long refactors and subtle bugs. For pure autocompletion, Claude Sonnet 4 (faster, more conservative) still wins, but for “fix this error I’ve been stuck on for 2 hours” GPT-5 is now the first choice.
More natural text
Less abuse of signal words (“fundamentally”, “it’s important to highlight”, “likewise”). The default style is a bit more direct, less academic. It still shows it’s AI if you leave it untouched, but less.
What stays the same (or gets worse)
- Speed: GPT-5 Thinking is slow. In tasks with deep reasoning, it can take 30-60 seconds. GPT-5 Instant compensates for fast chat.
- Memory: the “memory” function is still unreliable, it remembers weird things and forgets useful ones. Don’t rely on it for critical flows.
- Images: visual generation is still gpt-image-1, there’s been no generational leap. It’s still below Midjourney and Flux Pro.
- Voice: voice mode in Spanish works well, but there are occasional latencies and cuts that have been unsolved for months.
GPT-5 vs Claude Sonnet 4 vs Gemini 2.5 Pro in 2026
The synthetic benchmark counts for little. What matters is: what is each one better for today?
- Long analysis and editorial writing → Claude Sonnet 4 is still slightly ahead. Better with documents > 50 pages, better nuance in non-English languages.
- Programming and engineering → GPT-5 has taken the lead. Cursor and JetBrains already integrate it by default.
- Real-time search and voice → ChatGPT (with GPT-5) wins due to native integration with web, Canvas, and Tasks.
- Integration with Google Workspace → Gemini 2.5 Pro is still unbeatable if you live in Docs, Sheets, and Gmail.
Is it worth updating?
It depends on the plan you’re on:
- Free Plan: you already have GPT-5 Instant with limits. The update is free, it just changes the quality. Take advantage of it.
- Plus Plan (20 €/month): GPT-5 Thinking is included with generous limits. For medium professional use, this is where the sweet spot is.
- Pro Plan (200 €/month): only if you need GPT-5 Pro for very specific tasks (large dataset analysis, complex code). For most users, it’s excessive.
Typical mistakes when starting with GPT-5
- Expecting magic where there was work before. GPT-5 is better, not perfect. Creative tasks still require human editing.
- Letting it always choose Instant. For real reasoning, force Thinking explicitly with “think this through before responding”.
- Not updating your prompts. The prompts you wrote for GPT-4 may be too short. GPT-5 accepts richer context — take advantage of it.
Frequently Asked Questions about GPT-5
Is GPT-5 already available in Spanish?
Yes. Since the global launch, the model works in Spanish with the same quality as in English. The performance difference between languages has been drastically reduced compared to GPT-4.
How much does GPT-5 cost in the API?
The official prices in 2026 are approximately 1.25 € per million input tokens and 10 € per million output tokens for standard GPT-5. GPT-5 Pro and mini have different rates. Always check OpenAI’s official page before budgeting.
Does GPT-5 automatically replace GPT-4?
Yes. If you have a Plus or Pro plan, when you open ChatGPT, you’ll see GPT-5 as the default model. You can still use GPT-4 from the selector if you have old flows that depend on its exact behavior.
Is GPT-5 or Claude better for programming?
For refactoring and complex bugs, GPT-5 has taken the lead. For fast autocompletion in IDE and conservative code with good tooling, Claude Sonnet 4 is still very competitive. Many professional developers use both: Claude for continuous flow and GPT-5 for stuck problems.
Can GPT-5 access the internet in real-time?
Yes, ChatGPT with GPT-5 integrates native web search without needing to activate it manually. It cites sources and allows follow-ups. It’s more fluid than GPT-4’s separate search flow.
What this launch is for
The leap from GPT-4 to GPT-5 is not revolutionary but consolidating: less manual choice, better code, fewer hallucinations. If you used ChatGPT sporadically, you’ll notice the difference. If you use it daily, the improvement justifies the Plus investment on its own.
To start getting the most out of it without wasting time, check our guide to 10 ChatGPT prompts that save time at work — all compatible with GPT-5 — and if you’re unsure between assistants, the updated ChatGPT vs Claude comparison.
