The New Default. Your hub for building smart, fast, and sustainable AI software

See now
Illustration of ChatGPT's missing features and the future of OpenAI's GPT models

What Features Is GPT-4 Missing? 4 Limitations and How Newer GPT Models Changed Them

Matylda Chmielewska
|   Updated Nov 20, 2026

GPT-4, released by OpenAI on March 14, 2023, was missing four features: knowledge of recent events, long inputs, publicly available image input, and live access to online information. OpenAI’s launch page listed its knowledge cutoff as September 2021, its context length at 8,192 tokens, and called image input a research preview.

OpenAI launched ChatGPT, its chatbot, as a research preview on November 30, 2022, bringing generative AI to a mass audience. We asked Monterail engineers who use generative AI almost every day what they missed most in GPT-4, and this article pairs their wish list with what OpenAI has shipped since.

OpenAI’s newest models are GPT-6.2 Astra and GPT-6.2 Sol, as of November 2026. The sections below keep GPT-4 as the reference point and show how each limitation looks today.

Executive Summary

Every feature GPT-4 was missing at launch was a limit of one model release, and OpenAI's newest models now support all four. GPT-6.2 Astra lists a May 18, 2026 knowledge cutoff, a 1,050,000-token context window, image input, and web search as a supported tool. Two of our engineers' four wishes are now covered: EU rules address copyright, and open-source models allow self-hosting. Showing sources and catching misinformation still has to be built into each product. Product teams should pick the model that fits the job and keep it easy to swap, since limits change every few months.

What Is Generative AI?

Generative AI is machine learning that produces new text, code, images, or audio from patterns learned across many examples. For a broader introduction to tools such as ChatGPT, MidJourney, and Jasper, start with our guide to what generative AI is.

What Does GPT Mean in ChatGPT?

GPT stands for Generative Pre-trained Transformer. The name describes a model that generates text after learning from existing data, using the Transformer architecture, a neural network design for language tasks. GPT-4 is a large language model (LLM), and its limits are counted in tokens, the small chunks of text a model processes. ChatGPT is OpenAI’s chat app, built on GPT models such as GPT-4.

What Features Is GPT-4 Missing?

GPT-4 is a capable model with four functional limits. Knowledge stops at a training cutoff, and company data becomes available once a team connects it. Each prompt fits only a limited amount of text. At launch, image input and live data were not yet public. Newer models have changed this picture, and a later section of this article covers where GPT-6.2 Astra and GPT-6.2 Sol stand today.

What Does GPT-4 Know About Recent Events and Company Context?

GPT-4’s knowledge ends at its September 2021 data cutoff, so later websites and documents reach it only when you add them. The same goes for a company’s internal documents and expert knowledge: without that context, the output stays generic. Adding background to the prompt helps, and retrieval-augmented generation (RAG) goes further by searching your documents at question time and passing the relevant passages to the model.

Andreas Kollegger, Director of Applied AI Research at Neo4j, describes giving an agent company context as onboarding a new hire:

You onboard them. You tell them, “Here’s the company. Here’s the company culture. Here’s our customers. Here’s what they care about.”

Why AI Agents Need Knowledge Graphs | Andreas Kollegger, Neo4j (video)

How Much Text Can GPT-4 Handle at Once?

GPT-4 has a context length of 8,192 tokens, and a 32,768-token version (about 50 pages of text) was available in limited access as gpt-4-32k. That limit covered your prompt and the reply together. A long document had to be split into parts, and the longer the prompt, the shorter the reply could be.

Can GPT-4 Work With Images?

GPT-4 was built to take both images and text as input. It could work with document scans, photographs, diagrams, and screenshots, and it answered in text. At launch, though, image input was only a research preview, with Be My Eyes as OpenAI’s first partner, so the public could send text only. Image input reached ChatGPT Plus and Enterprise users in a rollout that began on September 25, 2023.

Can GPT-4 Access Real-Time or Location Data?

GPT-4 answers from its training data, so it cannot report current events or know where you are. OpenAI’s ChatGPT plugins, announced March 23, 2023, were the first fix. They let ChatGPT pull in online information and use third-party apps. Early access ran through a waitlist, with plugins from Expedia, Instacart, KAYAK, Klarna, OpenTable, Shopify, Slack, and Zapier. OpenAI has since deprecated plugins.

What Features Did Monterail’s AI Enthusiasts Want to See?

We asked four Monterail engineers and analysts who work with generative AI almost every day what they wanted next from ChatGPT, back when GPT-4 was still new. They wanted copyright safeguards, source transparency, misinformation detection, and private deployment. Read against today, their wish list shows how much has changed and what is still missing. The note under each quote gives the status as of November 2026.

Should AI Companies Disclose Their Training Data?

Krzysztof Kaiser, Head of Design & Business Analysis at Monterail, argues that AI vendors should disclose where their datasets come from and give the authors of that material an incentive to take part.

Many people voice their reasonable concerns regarding the security of AI tools, but there’s also the topic of copyright. While the texts, and images generated by artificial intelligence are for now not deemed copyrightable, we need a discussion around intellectual property and how to protect human-authored works of art, books, and scientific studies and ensure that companies such as OpenAI will have to disclose where their datasets are coming from and incentivize the authors of the materials used to build those similarly to how it’s been proposed in the early version of the recently released EU’s AI Act.

The EU has since turned part of this into law. Since August 2, 2025, providers of general-purpose AI models must draw up technical documentation, implement a copyright policy, and publish a summary of their model’s training content.

Should ChatGPT Show Its Sources?

Maciej Korolik, Senior Frontend Developer and AI Expert at Monterail, says that web access would reach all users and that showing sources would help people judge whether to trust an answer.

GPT plugins, web browsing, and search functionality are currently available for the ChatGPT Plus plan and a small group of developers, and they will be made available to the general public sooner or later. This will lead to the situation where ChatGPT’s ability to assess what information it should find online, and then add it to a response. If the chat would show the sources of information, it would be also easier to explain to someone why they should or should not trust the response they have received. I also believe that there will be more and more specialized AI-based tools where users will be able to find information i.e. only from scientific sources, with pre-made prompts.

Plugins did not become that route, because OpenAI deprecated them. Web search is now a built-in tool instead: the API documentation for GPT-6.2 Astra lists web search among its supported tools.

Can AI Detect Misinformation in Its Own Answers?

Darya Demidenko, a former business analyst at Monterail, asks for more transparency about the reasoning and sources behind generated answers, plus built-in detection of bias and misinformation.

I’d appreciate it if there was more transparency on the sources of generated insights and the reasoning behind them. I’d also like to see the ability to add specific domain knowledge and the customization of where the outputs may come from i.e. only backed up by specific scientific sources. Although ChatGPT may refuse to answer when asked about sensitive topics, it would be great if it could have in-built biases and misinformation detection feature to further secure the tool and ensure that it won’t be used for malicious purposes.

Both requests aim at answers a reader can check. Teams that need that today ground the model in vetted documents, which keeps answers tied to sources the team has approved.

Can Companies Run ChatGPT-Style Models on Their Own Servers?

Artur Rosa, Frontend Architect at Monterail, wants companies to build assistants that know their own products and to self-host models so sensitive data never leaves the network.

AI tools are already being used to assist in completing work-related tasks, so the natural thing here is using them to build an assistant that would know everything about a certain company or a product, and would help i.e. support these teams to perform more efficiently. One thing I'd really like to see, and something the AI community is also pushing towards, is the ability to self-host tools like ChatGPT and use them locally without the need for internet access. This would allow us to use the model for sensitive internal data as well and would address the security concerns that people have about using AI and uploading their data to external servers.

Open-source LLMs now make that possible: they can be deployed on private infrastructure so sensitive data stays in-house, as our guide to open-source LLMs explains. Whether on-premise hosting costs less than cloud depends on utilization, because owned hardware tends to win only for sustained, high-volume workloads.

Maxime Labonne, Head of Post-Training at Liquid AI, lists strict privacy requirements among the cases where a cloud model is not an option:

If you have strong requirements in terms of privacy, and it’s the case with a lot of regulated industries, then it’s also not really an option. Even though there are some ways to get around it, it’s never the top level of confidentiality.

Why Build AI Models 1,000x Smaller? | Maxime Labonne, Liquid AI (video)

When Was Each GPT Model Released? OpenAI’s Timeline From GPT-1 to GPT-6.2

OpenAI’s CEO, Sam Altman, was president of Y Combinator from 2014 to 2019 before becoming CEO of OpenAI. The timeline below lists the main releases, from the first GPT model to the newest.

  • December 11, 2015: OpenAI is introduced as a non-profit AI research company.

  • June 11, 2018: OpenAI publishes the first GPT model, a language model trained on a few thousand books, about 5 GB of text.

  • February 14, 2019: GPT-2 is announced with 1.5 billion parameters, trained on 40 GB of internet text from 8 million web pages. OpenAI did not release the full model at the time.

  • May 28, 2020: The GPT-3 paper describes a model with 175 billion parameters, and the OpenAI API follows on June 11, 2020.

  • November 30, 2022: ChatGPT launches as a free research preview.

  • March 14, 2023: OpenAI releases GPT-4.

  • March 23, 2023: ChatGPT plugins are announced; OpenAI has since deprecated them.

  • April 30, 2025: GPT-4 leaves ChatGPT, where GPT-4o replaces it, and API access continues.

  • August 7, 2025: OpenAI introduces GPT-5.

  • September 3, 2026: OpenAI unveils GPT-6 Astra.

  • November 17, 2026: GPT-6.2 Astra and GPT-6.2 Sol launch in ChatGPT Work, Codex, and the API.

The wait between flagship generations narrowed from about 29 months (GPT-4 to GPT-5) to about 13 months (GPT-5 to GPT-6).

Where Does ChatGPT Stand Today, and What Comes Next?

What Is OpenAI’s Latest Model Today?

OpenAI’s latest models are GPT-6.2 Astra and GPT-6.2 Sol, released on November 17, 2026. Astra is the frontier model for the hardest projects, and Sol is the efficient model for everyday work. Both are available to ChatGPT Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and to developers through the API. OpenAI highlights multistep agentic workflows and computer use as strengths, and says Astra produces slides, spreadsheets, and documents that follow your templates and the conventions of your field.

GPT-4 itself has left ChatGPT. OpenAI removed GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT on February 13, 2026. OpenAI’s announcement benchmarks GPT-6.2 Astra against Anthropic’s Claude Fable 5.2.

Which of GPT-4’s Limitations Do Newer Models Address?

The table sets GPT-4’s launch figures next to GPT-6.2 Astra’s published specifications.

Limitation

GPT-4 (March 2023)

GPT-6.2 Astra (November 2026)

Knowledge cutoff

September 2021

May 18, 2026

Context length

8,192 tokens (32,768 in gpt-4-32k)

1,050,000 tokens

Image input

Research preview, not publicly available

Text and image input, text output; image generation available as an API tool

Live and company data

No built-in tools; plugins by waitlist

Web search, file search, computer use, and Model Context Protocol (MCP) tools in the API

Each capability still needs setup. The cutoff applies, so anything newer than May 18, 2026 needs web search or supplied documents. Companies that start with one measurable generative AI process can add those tools step by step, and private hosting keeps prompts inside the company network instead of passing redacted versions to outside vendors.

What Comes Next for ChatGPT?

Release pace is the best guide available. OpenAI shipped GPT-6.2 about two and a half months after GPT-6 Astra, and the models in between change what teams can build: Monterail tested GPT-4.1 against GPT-4o-mini in a deployment-style workflow, for example.

A design that treats LLM integration as a swappable layer lets a team adopt a newer model without a rebuild. Our guide to deciding whether to build, buy, or outsource AI recommends a hybrid model that buys commodity tools and builds proprietary decision logic.

Key Takeaways

  • GPT-4 launched on March 14, 2023 with a September 2021 knowledge cutoff and an 8,192-token context length; image input was a research preview.

  • GPT-6.2 Astra, released November 17, 2026, lists a May 18, 2026 cutoff and a 1,050,000-token context window, and it accepts image input and supports web search as a tool.

  • ChatGPT plugins, the 2023 answer to live data, have been deprecated, so web search and file search now supply live and company data.

  • Since August 2, 2025, EU rules require providers of general-purpose AI models to publish a summary of their training content and keep a copyright policy.

  • Source transparency and misinformation checks depend on how a team builds around the model, so private hosting and swappable models are design choices worth making early.

How Should Product Teams Plan Around Changing GPT Models?

GPT-4’s four limitations came from one model release, and newer models list each of them differently. A product team gets the most from AI by matching the model to the job and keeping the architecture flexible as models change every few months.

A custom build adds your own data and controls on top of whichever model leads. If you want to develop your own generative AI solution tailored to your product, explore our AI development services.

GPT-4 Features and ChatGPT Models FAQ

Matylda Chmielewska avatar
Matylda Chmielewska
Former Content Specialist
Linkedin
Experienced marketing specialist specialized in the tech sector. Currently works as a Junior Carbon Specialist at a global corporate design, engineering and sustainability company, where she helps large and medium-sized companies make their first steps towards a more sustainable business model by providing ESG reporting services and sustainable development strategies.