What can ChatGPT really do?, 2026 status

Three years after the public launch, ChatGPT is no longer hype for most people but a tool they either use every day or not at all. The question “is this any good?” has been answered. The question “good at what, exactly, and how reliably?” has not.

Here is an honest stocktake of what works dependably in 2026, where it still wobbles, and what has changed since 2024.

What works dependably

Language & text

• Translation between major languages: excellent, often better than DeepL when it comes to understanding context
• Summarizing texts up to the context-length limit: reliable
• Rephrasing / adjusting tone: very good
• Proofreading: solid; but it misses stylistic inconsistencies
• Brainstorming: excellent for the first 20 ideas, repetitive after that

Structuring & researching

• Lists, tables, overviews from unstructured text: reliable
• Explaining complex concepts at different levels: solid, with caution on the details
• Comparison tables from multiple sources: good, when the sources are provided alongside

Code

• Generating boilerplate: reliable
• Explaining bugs: very good
• Suggesting refactorings: solid
• Writing complete apps: only with a clear specification and code review

Multimodal (since 2024)

• Describing image content: excellent
• Reading diagrams: good, with gaps on very dense charts
• Voice input: reliable even in dialect

What wobbles

Current facts

ChatGPT hallucinates. On dates, figures, studies, sources. The 2026 versions with web search reduce this sharply, but only when web mode is active and the search finds something. With current events or niche topics: still proceed with caution.

Math & logic with many steps

Once you chain about 5 steps together, it gets shaky. The o-series (reasoning models) has improved this, but for business-critical calculations: not without human review.

Keeping long texts consistent

Keep the style consistent across a 30-page document? Difficult. Models “forget” earlier decisions, especially once the context window fills up.

Creative writing

Usable for first drafts. But: typical ChatGPT turns of phrase (“In conclusion”, “It’s important to note”) are recognizable, and the tone comes across as smoothed over.

Source citation

Even with RAG: not 100% reliable. Sources are occasionally invented or wrongly attributed.

What’s new in 2026

Longer context windows. Most top models now have 200,000+ tokens. You can paste in entire books and have them queried, without hallucinations about content that is no longer in the window.

Reasoning models. The o-series (OpenAI), Claude 4 (Anthropic), Gemini Deep Think, all of them “think longer” before they answer. Accuracy on math, logic, and multi-step tasks is noticeably better. The cost: higher latency and higher token prices.

Agent capability. Models now carry out multi-step tasks: “Book me an appointment in the calendar app and send the person a confirmation.” Still experimental in 2024, in 2026 it is production-ready for limited domains.

Multimodal as standard. Image + text + (sometimes) audio in a single conversation. You photograph a table, the model extracts it and responds to it.

Open source is catching up. Llama 4, DeepSeek, Qwen, models you can host yourself, are no longer significantly worse in 2026 than the top closed models.

Where it isn’t yet

• Real creativity, ChatGPT combines what already exists in new ways. Genuine conceptual leaps: rare.
• Self-criticism, when the model is wrong, it often doesn’t know it.
• Goals of its own, agents have goals, but only the ones you give them.
• Domain depth, a model is a generalist. Specialist knowledge needs fine-tuning or RAG with expert sources.

The honest recommendation

Try three tasks from your working day:

1. One you can easily imagine it doing (drafting an email reply)
2. One you don’t trust it to do (extracting tables from PDFs)
3. One that’s meant to be creative (writing ad copy)

Compare the result with your own output. These three experiments say more than 1,000 marketing demos.

---

In 2026, ChatGPT is a tool on the level of a good intern: enthusiastic, fast, with gaps. Those who understand that get more out of it.