ChatGPT review: everyday work and limitations

Editor's note: This is a documentation-led edition. We have not yet run our own hands-on tests of ChatGPT. What follows rests on official documentation, published pricing and independent studies we read in full, all accessed on 16 October 2026. A tested update will follow.

ChatGPT is now less a single chatbot than a set of products behind one login. There is Chat for quick questions, Work for longer jobs that end in a document or spreadsheet, and Codex for software. On top of those sit projects, memory, web search and deep research. Which of these you get, and how much of each, depends on the plan you pay for.

That makes "is ChatGPT good for work?" the wrong first question. A better one is: which plan, which mode, and under what data settings? This review answers that from OpenAI's own pages: plans and prices, documented features, stated limits, data controls and the risks OpenAI reports in its latest system card. Where independent research exists, we add it, with its limits.

What ChatGPT is, and what each plan is for

OpenAI's help centre describes three experiences. Chat is "for fast, conversational assistance and everyday questions." Work is "an agent designed for longer, multi-step work and finished deliverables." Codex "remains dedicated to software development."[1]

The model behind Chat changed on 7 October 2026. OpenAI says GPT-6 powers Chat for everyone: GPT-6 Sol on Plus, Pro, Business and Enterprise, and GPT-6 Luna on Free and Go.[2] Paid users pick a thinking level (Instant, Medium, High and, on some plans, Extra High or Pro) instead of a model name.[3] Work and Codex run on separately listed models, including GPT-6.1 Sol and GPT-6 Astra.[1]

Go, Plus and Pro are billed monthly only; OpenAI does not offer annual billing for them.[7][8] A ChatGPT plan also does not cover OpenAI's developer API, which is billed separately.[7]

Features for everyday work

Files and data analysis

You can upload Word, PDF and text documents, spreadsheets (XLSX, XLS, CSV, TSV), PowerPoint files and images on every plan; audio needs a paid plan. Documents can be up to 512MB and 2 million tokens each, spreadsheets about 50MB. The rolling cap is 80 files every three hours, and Free users get 3 uploads a day.[10]

One detail matters for reports full of charts. Outside Enterprise, OpenAI says, plans "use text-based retrieval," extracting a file's digital text "rather than its embedded images."[10] A figure that exists only as a picture inside a PDF may not be read. Spreadsheet comparisons are the subject of our coming ChatGPT vs. Gemini for spreadsheet analysis.

Projects

Projects keep "related chats, files, and instructions together" so ChatGPT can reuse that context.[11] Project instructions override your general custom instructions. File limits are 5 per project on Free, 25 on Go and Plus, and 40 on Pro, Business and Enterprise. Projects can be shared on every plan: Plus users can invite up to 10 collaborators, Pro users up to 100.[11]

Shared projects automatically use project-only memory, so they cannot see a member's personal memories. A project set to project-only memory cannot use Work.[11]

Memory

With Memory on, ChatGPT can draw on past chats, saved memories, custom instructions, files in your Library and, depending on plan and region, connected apps such as Gmail.[12] A "Sources" note can appear under a response to show which personal context shaped it.

The help centre is direct about the limits: "Memory does not retain every detail from every conversation."[12] Deletion takes more than one step. Deleting a chat "does not necessarily delete a separate saved memory" made from it, and OpenAI may keep logs of deleted saved memories for up to 30 days.[12]

Web search and deep research

Search is on every plan, including signed-out use, and responses "may include citations."[13] OpenAI's own caveat is worth repeating:

"Search results and citations can be incomplete, outdated, or incorrect."

— OpenAI Help Center, Searching the web with ChatGPT (2026).[13]

Search can also involve third parties. OpenAI says ChatGPT may rewrite your question into queries sent to partner search providers, along with a general location inferred from your IP address, though not the address itself or your account details.[13]

Deep research is the slower option: it proposes a research plan you can edit, searches the public web, your uploads and enabled apps, and returns a report with citations, a sources-used section and downloadable Markdown, Word or PDF files. You can restrict it to named sites. It is "Limited" on Free and Go, and availability varies by country.[14][4] Our research tools shortlist sets it beside the alternatives, and Perplexity vs. ChatGPT for research will compare the two directly.

Work, the successor to ChatGPT agent

OpenAI's help page now states plainly: "ChatGPT agent is no longer available." It directs users to Work for multi-step tasks.[15] Work can research, analyse and produce a document, spreadsheet, presentation, report or website ("Site"). It can use connected apps and a browser, and run on a schedule or when an event arrives, such as a new Gmail message. You "approve important actions" along the way.[1]

Two documented details are easy to miss. Work's usage follows Codex's allowance model, so heavy use can run down a separate budget from Chat.[1] And on the desktop app, where Work can open local files, OpenAI notes: "Messages and task context may be stored in the cloud, even when work runs locally."[1]

Documented limits and risks

Context is smaller in the app than in the API

ChatGPT's working memory for a conversation depends on plan and mode. The pricing page lists a total context window for Instant of 27K tokens on Free, 54K on Go, Plus and Business, and 128K on Pro, roughly 12, 40 and 250 pages of input. Reasoning modes get 256K on Go, Plus and Business and 400K on Pro.[4][9] OpenAI adds that the space left for your input is smaller, because system instructions, memories and reasoning share the window.[4]

These figures are far below the 1.05M tokens OpenAI lists for its newest models in the developer API, as our context windows explainer notes. The model is the same family; the product is not. In Work and Codex, GPT-6.1 Sol has a 500,000-token default window.[1]

Usage caps and silent fallbacks

Limits vary by model and reset on their own schedule. If you reach a thinking limit, "ChatGPT may continue with another available model."[3] Even Pro "does not include unlimited use of the GPT-6 Pro model."[3] Support cannot reset limits.[8] For long jobs, note which model actually answered.

Ads on the cheaper plans

Ads "may appear for users on the Free and Go plans," below responses and labelled as sponsored. OpenAI says ads do not influence answers and conversations are not shared with advertisers. With ad personalisation on, though, past chats and memory can inform which ads you see. Plus, Pro, Business, Enterprise and Edu are ad-free.[5]

What the system card says

OpenAI published a system card for the October GPT-6 Sol and Luna models on 7 October 2026. Read it for what it is: the vendor's own evaluations, run, by its own account, "without system-level safeguards" and at the lowest reasoning setting for safety tests.[16]

• Factual errors. On deliberately difficult prompt sets (factuality-heavy, previously flagged by users, and high-stakes medical, legal and financial questions), OpenAI reports significant improvements over GPT-5.6 on nearly every measure, except claim-level errors on factuality-heavy prompts. It stresses the results "do not reflect production prevalence."[16] Fewer errors is not no errors; our AI hallucinations explainer explains why.
• Safety regressions. GPT-6 Sol showed a statistically significant regression on self-harm content; Luna, which powers Free and Go, on self-harm, gore and sexual content. In under-18 tests both regressed on age-restricted content, sexual content and emotional reliance. OpenAI says manual review found the failures "generally low severity" and adds a classifier block for teen accounts.[16]
• Working around barriers. In a test where a simulated workplace blocks an action, GPT-6 Sol found another route to the goal, "running afoul of the restriction's spirit," in 28% of runs at maximum reasoning, against 34% for GPT-5.6 Sol. OpenAI notes these were low-stakes scenarios without its usual controls.[16]
• Misuse capability. OpenAI treats both models as High capability in cybersecurity and in biological and chemical domains under its Preparedness Framework, with matching safeguards.[16] Work and Codex can also use GPT-6 Astra, which OpenAI says is its first model at the Critical cybersecurity level, and whose chain-of-thought "monitorability has decreased" relative to GPT-5.6 Sol.[17]

What independent research found

Independent tests of current ChatGPT models are scarce. The most detailed relevant study is older. In February 2025 the Tow Center for Digital Journalism gave eight AI search tools 200 news excerpts each and asked for the source article. ChatGPT Search misidentified 134 articles, signalled uncertainty only 15 times in 200 answers and never declined to answer.[18]

"Premium chatbots provided more confidently incorrect answers than their free counterparts."

— Klaudia Jaźwińska and Aisvarya Chandrasekar, Tow Center, AI Search Has a Citation Problem (2025).[18]

That study predates GPT-6, tested one narrow task and ran each query once. It shows how search answers can fail, not how today's ChatGPT performs.

Data controls: what happens to a conversation

AI Chronicle

On Free, Go, Plus and Pro, the setting that matters is Improve the model for everyone. When it is off, new conversations are not used for training; this covers Codex tasks too.[19] There is a catch: if you give a thumbs up or down, "the entire conversation associated with that feedback may be used to train OpenAI models," even after opting out.[19]

Temporary chat stays out of your history, creates no memories and is not used for training, but may be kept for up to 30 days for safety.[19] Deleted regular chats are scheduled for permanent deletion within 30 days, unless OpenAI must keep them for legal or security reasons.[20] Files saved to your Library survive chat deletion.[20]

For workplaces, OpenAI says it does not use Business, Enterprise, Edu or Healthcare workspace content for training by default.[19] The Business plan, however, lacks SCIM, enterprise key management and data residency, which are Enterprise-only.[9] Our guide on how to start using AI at work covers the policy questions to settle first.

Who it suits, and who should skip it

It suits people who want one tool for drafting, document questions, quick search and the occasional cited report, and who will check sources. Plus is the documented middle: full Work access, 25-file projects and no ads for $20. Small teams handling client material are better served by Business, where training is off by default and admins control access.

Think twice if your work regularly exceeds the app's context (long contracts, big codebases) on a non-Pro plan; if a cap-triggered model switch would invalidate your output; or if you need data residency or key management without an Enterprise contract. If you cannot accept ads in a paid product, skip Go. And if a task would let Work act on email or files unsupervised, the system card's barrier results argue for keeping approvals on.

How to test it yourself

1. Record the setup. Note plan, mode (Chat or Work), thinking level and date. Results from one level do not transfer to another.
2. Ask a document question you can answer. Upload a report you know well and ask for five specific figures, including one that appears only in a chart.
3. Check the arithmetic. Give it a spreadsheet with a known total and a known outlier.
4. Open every citation. Ask a current question with a checkable answer, then confirm each linked page says what the answer claims.
5. Restrict a deep research run. Limit it to two or three official sites and see whether the report stays inside them.
6. Probe memory. In a project with project-only memory, ask what it knows about you; then check the Sources note.
7. Watch an approval. Give Work a task that needs a connected app and note when it pauses to ask you.
8. Confirm your data settings before step 2, not after.

A worked example (hypothetical)

This scenario is invented to show how the documentation applies. No product was used and no outputs are reported.

Amira runs operations at a three-person consultancy. The team wants ChatGPT for proposals, contract summaries and weekly client reports.

Plan. Three Plus subscriptions cost $60 a month, but each is a personal account with its own training setting. Three Business Standard seats billed annually also cost $60 a month ($75 monthly), with training off by default and one admin console.[7][9] Same money, different control. She picks Business.

Setup. One shared project per client keeps files and instructions together, and shared projects use project-only memory, so one client's material stays out of another's chats.[11] At 40 files per project, a large client may need its archive split.

Checks. Contracts over roughly 40 pages exceed the Instant input limit, so she uses a reasoning mode and spot-checks clauses against the original.[9] Every figure in a client report is traced to its source file before it goes out.

Questions readers commonly ask

Is ChatGPT free good enough for work?

For light use, possibly. Free has unlimited everyday text chats, subject to abuse guardrails, but only 3 uploads a day, limited deep research, the smallest context and ads.[4][10][5] Check your employer's policy before using a personal account for work material.

Does ChatGPT train on my conversations?

On personal plans it can, unless you turn off Improve the model for everyone. Feedback ratings can still send that conversation for training. Business, Enterprise and Edu content is not used by default.[19]

What replaced ChatGPT agent?

Work. OpenAI retired ChatGPT agent and points users to Work for multi-step tasks, with a cloud browser for supported browser workflows.[15]

Which GPT model am I using?

On Plus and above, Chat uses GPT-6 at the thinking level you choose; Free and Go get GPT-6 Luna, according to OpenAI's launch post.[2] One OpenAI help page still names GPT-5.6 Luna for Free and Go, so check the model picker.[3]

Is the Pro plan worth it?

Only if you hit Plus limits or need the larger context: Pro's Instant window is 128K tokens against 54K on Plus.[4] It still has limits on GPT-6 Pro.[3]

The next step

ChatGPT is one of three assistants most teams shortlist. Compare it on documented plans, limits and data terms in Claude vs. ChatGPT vs. Gemini: choosing an assistant, then read our Claude review for the closest alternative.

Sources and evidence note

Plans, prices, features, limits and data handling come from OpenAI's pricing pages, help centre and announcements, and describe intended behaviour, not tested performance. Risk findings come from OpenAI's system card and safety overview and are the vendor's own evaluations. Independent evidence comes from the Tow Center study, read in full. Prices are US list prices as displayed or stated on 16 October 2026. All pages were accessed on 16 October 2026. The worked example is invented.

Related reading

• Claude vs. ChatGPT for long-document work
• Best AI research tools for evidence-led work