ChatGPT – the product, capabilities, and limits

ChatGPT is OpenAI’s consumer-facing conversational AI product, distinct from the underlying GPT models and from the developer-facing OpenAI API. As of May 2026, it serves roughly 900 million weekly active users, runs GPT-5.5 as the flagship model on paid tiers, and ships across six subscription tiers from a free plan with ads to a $200-per-month Pro plan with a one-million-token context window.

That definition is accurate and not very useful. The harder question, the one you came here for if you build voice agents, avatars, or any customer-facing digital-human system, is which slice of ChatGPT to embed and where it stops being the right tool. The product is one option; the OpenAI API is another; Azure OpenAI Service is a third; an entirely different vendor is a fourth. This article maps the capabilities, the honest limits, and the strategic choice between them.

Everything below is dated to May 2026. OpenAI changes pricing, model names, and feature gates roughly every six weeks. Verify the specific numbers against openai.com/pricing before you commit to a budget.

What ChatGPT is in 2026

Three things often get conflated and should not be. ChatGPT is the product, the chat interface at chatgpt.com and the mobile apps. The OpenAI API is the developer endpoint at api.openai.com, billed per token. GPT-5.5, GPT-5.4, GPT-5.3 and the older GPT family are the underlying models. The product wraps the model; the API exposes it. Pricing, retention defaults, and feature availability differ across all three.

The product currently runs GPT-5.5 (internal codename Spud), released on April 23, 2026, as the default model on Plus, Pro, Business, and Enterprise. Free users get GPT-5.3. GPT-5.5 Pro is reserved for the $100 and $200 Pro tiers and Enterprise. The older models that powered ChatGPT through 2024 and 2025, including GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, and the original GPT-5 Instant and Thinking variants, were retired from the product on February 13, 2026, with limited Custom-GPT availability through April 3. API access to some retired models continues separately.

The numbers around the product are large enough to distort intuition. Roughly 900 million weekly active users; 50 million paying subscribers across Plus, Go, Pro, Business and Enterprise; about $25 billion in annualized revenue as of February 2026; and 92% of the Fortune 500 with at least some ChatGPT deployment. OpenAI processes around two billion prompts per day on the consumer product alone.

The corporate structure changed in late 2025 in ways that matter for procurement and legal review. OpenAI completed its recapitalization on October 28, 2025, creating OpenAI Group PBC, a public benefit corporation operationally controlled by the OpenAI Foundation (the renamed nonprofit). Microsoft holds approximately 27% of OpenAI Group PBC, valued at $135 billion, with IP rights extended through 2032 and a $250 billion Azure commitment from OpenAI as part of the deal. If your compliance review touches ownership chains or counterparty risk, the PBC structure is the current answer.

The capabilities map (the honest version)

ChatGPT does many things. Most articles list them in marketing order. The order below is what actually matters when you are deciding whether to ship it.

Text generation and reasoning. GPT-5.5 scores 92.4% on MMLU and 88.7% on SWE-Bench Verified, with roughly 60% fewer hallucinations than GPT-5.4 on OpenAI’s internal evaluations. These numbers put it at the top of the frontier alongside Claude Opus 4.7 and Gemini 3.1 Pro. They do not put it at human-expert level on novel problems, and a 60% reduction in hallucinations is not the same as zero. For full coverage of the model family, see our GPT acronym and architecture deep dive and the GPT-4 capabilities deep dive for the prior generation.

Code generation. Strong on isolated functions and on real GitHub issues at the benchmark level. 82.7% on Terminal-Bench 2.0, 88.7% on SWE-Bench Verified, and roughly 40% fewer output tokens than GPT-5.4 on equivalent Codex tasks. Builders shipping production code should treat the benchmark scores as a ceiling, not a floor, and budget for review.

Voice. Voice Mode in ChatGPT runs on the same family but with a distinct latency and audio pipeline. We cover this separately in ChatGPT Voice Mode explained because the engineering trade-offs around interruption, turn-taking, and barge-in deserve their own analysis.

Image generation. ChatGPT uses gpt-image-2, shipped on April 21, 2026, as the default image model, replacing GPT Image 1.5. Quality is competitive with Stable Diffusion 3 and Imagen at the level a non-specialist will notice. We unpack the DALL-E lineage and integration trade-offs in our ChatGPT image generator deep dive.

File analysis. PDFs, spreadsheets, code, and most common document types. Context window matters here. Plus and Business users get the standard 256K-token context; Pro $200 unlocks the 1M-token context window. Past those caps, the model loses coherence on cross-document reasoning rather than failing cleanly.

Web browsing and tool use. Live web access through the integrated browser; computer use through Operator on Pro tiers and above; the GPT Store and Custom GPTs for shareable task-specific assistants. Codex Mobile shipped free on every plan, including Free and Go, on May 14, 2026.

Long-horizon agentic work. GPT-5.5 was positioned by OpenAI as the first model designed for what they call “the next way of getting work done on a computer”, with internal benchmarks on tasks of 20-hour median completion time. The framing is ambitious; the production reality is that agentic workflows still need supervision and fallback paths. We cover the framework choices in detail in our Agentic AI silo, starting with the canonical pillars there.

What ChatGPT cannot do (the limit map)

The capability list is well-documented. The limits are not, and that asymmetry is where bad architectural decisions get made.

Real-time knowledge. GPT-5.5 has a December 2025 training cutoff. Anything that happened after that exists for the model only if a tool call retrieves it. If your product asks ChatGPT a question about a news event from last week and you have not given it a browsing or search tool, it will not say “I don’t know”; it will produce a plausible-sounding answer that may or may not be accurate.

Specialized factual accuracy. Medical, legal, and financial questions remain a known weakness, even on the frontier models. A December 2025 peer-reviewed benchmark of GPT-5 on the American College of Radiology in-training exam, published in Frontiers in Oncology, found measurable gains over prior models but persistent need for expert oversight on case-level recommendations. Treat domain-specialized output as a draft that requires a credentialed human, not as an answer.

Long-document coherent reasoning beyond the context window. The 1M-token context on Pro $200 is real but expensive to use well. Retrieval-augmented generation usually outperforms naive long-context use on cost-per-quality, which is why we cover RAG and vector infrastructure as a separate concern in the architecture stack.

Reliable tool use without explicit guardrails. Function calling works. It works without producing the calls you want when the prompt is loose. Production systems wrap tool calls in validation layers; ChatGPT the product does not give you that wrapper.

Compliance-critical deterministic workflows. Tax filing, regulatory submissions, anything where the output must match a spec exactly. Probabilistic models are the wrong layer for these. Use ChatGPT to draft, not to file.

Honest reasoning about its own limits. The model will tell you it cannot do something when prompted directly, and it will then attempt to do that thing if the next prompt frames the request differently. Asymmetric confidence in the prompt yields asymmetric output. This is not a bug you can prompt-engineer away; it is a property of the architecture. For a deeper analysis, see How accurate is ChatGPT.

ChatGPT vs the OpenAI API: the strategic choice for builders

This is the section that decides what you actually ship. The choice is not subtle.

The product (ChatGPT Plus, Business, Pro). You get a polished chat UI, voice mode, Custom GPTs, file upload, the image and video integrations, Deep Research, and a flat monthly fee. You do not get programmable access, you do not control system prompts beyond what Custom GPTs allow, and you cannot embed any of it in a customer-facing application except by referring your users to chatgpt.com.

The API. You get programmable access to the same model family, full control of system prompts, function calling, structured outputs, fine-tuning, embeddings, and the ability to embed responses anywhere. You pay per token. You build the UI. You handle retries, rate limits, error states, and observability yourself.

The cost arithmetic for a single user looks like this, as of May 24, 2026. ChatGPT Plus is $20 per month with soft caps that most non-power users will not hit. GPT-5.5 in the API is $5 per million input tokens and $30 per million output tokens, double the GPT-5.4 rate. A moderately heavy API user generating 500,000 output tokens per month pays $15 for output, plus input cost, before storage, caching, and engineering time. Below that volume, Plus wins on raw cost; above it, the API wins on raw cost but loses on the engineering bill. Cached input gets up to a 90% discount on repeated context; the Batch API gives 50% off with a 24-hour async turnaround. We cover the full pricing matrix in ChatGPT API authentication and pricing and the consumer-tier value analysis in ChatGPT Plus pricing.

For builders of digital-human products, the rule of thumb is simple. If your team needs to draft, research, and ideate, use ChatGPT Business. If you are embedding conversational AI in a product your customers see, use the API. The product is not licensed for embedding. The API is.

The customization gap matters more than most decision-makers realize. Custom GPTs let you set a system prompt, attach files, and define actions. The API lets you control every token, version the system prompt, A/B test variants in production, and switch models without rebuilding the user experience. If your roadmap includes anything beyond a single canonical assistant, the API is the only path.

How ChatGPT fits into OpenAI’s product strategy

ChatGPT is the front door. Most of what OpenAI ships in 2026 is downstream of it or designed to feed traffic into it.

The model lineage runs through the GPT family for general use and the o-series for reasoning, with the o-series now folded into GPT-5.5 rather than maintained as a fully separate product. See our OpenAI o-series lineage for the history and the current state. Sora handles video generation; Codex handles coding and computer-use agents; ChatGPT Atlas is OpenAI’s experimental browser; gpt-image-2 covers image generation inside ChatGPT and via API.

The Microsoft partnership remains the backbone of distribution and infrastructure. Microsoft keeps exclusive IP rights to OpenAI’s models and Azure API exclusivity through 2032. We unpack the dependency in detail in the Microsoft–OpenAI partnership analysis. The short version for builders: if you are on the Microsoft stack already, Azure OpenAI Service is often the lower-friction path to the same models, with European data residency and enterprise contracting handled at the Azure layer.

OpenAI’s stated thesis is platform consolidation: ChatGPT becomes the daily-use AI, the API becomes the developer substrate, and Sora, Codex, Atlas, and the gpt-image family slot in as features rather than separate products. Whether you accept that thesis or hedge against it shapes your dependency choices. The full company-strategy analysis lives in our OpenAI company and ecosystem pillar.

Safety, privacy, and what OpenAI does with your conversations

This is the section that determines whether your security team will let you ship.

The defaults differ by tier. On Free, Plus, and personal Pro accounts, your conversations may be used to train future OpenAI models unless you turn off “Improve the model for everyone” in Data Controls. The opt-out applies to new conversations only; anything sent before the toggle is set is already past the gate. Memory, the long-term context feature, is stored separately and persists until you clear it manually. Temporary Chat is the cleanest consumer-level option: not stored, not used for training, no memory.

On Business, Enterprise, and Edu, conversations are not used for training by default and no opt-out is required. Workspace admins control retention. On the API, inputs and outputs are retained for 30 days for abuse monitoring and then deleted, and they are not used for training. Zero Data Retention (ZDR), which eliminates the 30-day window entirely, is available only to qualifying enterprise customers who apply through their API account.

Two non-obvious points matter for compliance review. First, even after you delete a chat, OpenAI retains it for 30 days for abuse monitoring; the data window exists outside your control during that period. Second, an ongoing court order from US litigation has required OpenAI to preserve certain deleted ChatGPT logs beyond the standard 30 days. If you are in a regulated industry, the API with ZDR or a different vendor entirely is the only configuration that holds. We cover the security review in detail at Is ChatGPT safe and the privacy-specific analysis at Is ChatGPT private.

OpenAI’s data controls documentation is at help.openai.com. Read the page that applies to your tier; the consumer defaults and the API defaults are genuinely different.

The bottom line for builders

Three decisions cover most of the field.

If your team needs daily AI for drafting, research, and ideation, and your conversations are not regulated, ChatGPT Business at $20 per seat per month with a two-seat minimum is the standard answer. SOC 2, SAML SSO, no training on your data, and a credible enough security posture for most internal use. Plus at $20 a month works for a single power user; Pro $100 or Pro $200 only makes sense if you actually need Sora, the 1M-token context, or near-unlimited Deep Research.

If you are embedding conversational AI in a product your customers see, the API is the answer. Plus and the consumer plans are not licensed for embedding, and using them by piping prompts through a browser session is a violation of the terms that will get your account closed at scale. Build on the API directly, on Azure OpenAI Service if your stack is Microsoft, or on a different model family if your evaluation favors Claude, Gemini, or open-source. The full comparison set sits in our Vendors and Models hub.

If you are in a regulated industry (healthcare, finance, legal, government), the only configuration that holds under audit is API access with ZDR, or a model deployment where the data path is fully under your control (Azure OpenAI in your tenant, a private deployment, or an open-weights model running on your infrastructure). The consumer product is not the right layer regardless of how convenient it is.

Frequently asked questions

What is ChatGPT?

ChatGPT is OpenAI’s conversational AI product, available as a web app at chatgpt.com and as mobile apps. It runs on the GPT model family (currently GPT-5.5 on paid tiers, GPT-5.3 on Free) and is separate from the OpenAI API, which gives developers programmable access to the same underlying models.

Is ChatGPT free?

Yes, with limits. The Free tier gives access to GPT-5.3 with a cap of 10 messages every five hours and shows ads in the US (since February 9, 2026). For unlimited GPT-5.5 access without ads, Plus at $20 per month is the entry-level paid plan.

Which model does ChatGPT use in 2026?

As of May 24, 2026, ChatGPT defaults to GPT-5.5 (released April 23, 2026) on Plus, Pro, Business, and Enterprise plans. Pro tiers and Enterprise also get GPT-5.5 Pro. The Free tier uses GPT-5.3. Older models including GPT-4o and GPT-5 Instant were retired from the product on February 13, 2026.

How much does ChatGPT Plus cost?

ChatGPT Plus is $20 per month and has held that price for three years. OpenAI has publicly suggested the price could rise to roughly $44 per month by 2029. Pro $100 and Pro $200 are the higher tiers; Business is $20 per user per month with a two-user minimum.

What is the difference between ChatGPT and the OpenAI API?

ChatGPT is the chat product, billed monthly, not licensed for embedding in your own application. The OpenAI API is the developer endpoint, billed per token, with full programmatic control and explicit permission to embed. Builders shipping customer-facing AI use the API; teams using AI internally typically use ChatGPT Business.

Does ChatGPT train on my conversations?

On Free, Plus, and personal Pro accounts, yes, unless you opt out in Settings > Data Controls > Improve the model for everyone. On Business, Enterprise, Edu, and the API, conversations are not used for training by default and no opt-out is required.

How long does ChatGPT keep my data?

Conversations on consumer tiers are retained until you delete them, then for up to 30 days afterward for abuse monitoring. On the API, all inputs and outputs are retained for 30 days then deleted. Zero Data Retention (ZDR) is available only to qualifying enterprise customers. An ongoing US court order may require longer retention for certain logs.

Can ChatGPT browse the web?

Yes. The integrated browser tool is available on Plus and above, and Codex Mobile shipped free on every plan including Free and Go on May 14, 2026. The model still has a December 2025 knowledge cutoff for anything not retrieved through the browsing or search tools.

What is ChatGPT’s context window?

256K tokens on Plus and Business, 1M tokens on Pro $200 and Enterprise. The 1M-token context is real but expensive to use well; retrieval-augmented approaches typically outperform raw long-context for most production use cases.

Is ChatGPT better than Claude or Gemini?

Depends on the task. GPT-5.5 leads on coding (88.7% SWE-Bench Verified) and math (FrontierMath 4 at 35.4%). Claude Opus 4.7 leads on several SWE-bench variants and on enterprise coding tasks. Gemini 3.1 Pro leads on long-context retrieval and multimodal benchmarks. For production work, route different tasks to different models rather than committing to one.

Does ChatGPT have ads?

In the US, yes. Free and Go tiers have shown tinted-box ads alongside AI-generated answers since February 9, 2026, sold through OpenAI’s self-serve Ads Manager. Plus, Pro, Business, and Enterprise are ad-free.

What can ChatGPT not do?

It cannot reliably handle real-time knowledge without a browsing tool, specialized factual accuracy in medical, legal, or financial domains, deterministic compliance-critical workflows, long-document reasoning beyond its context window, or honest reflection on its own limits when the prompt is adversarial. Treat output as a draft requiring review.

Should I use ChatGPT for my business?

For internal team productivity, yes, on Business at $20 per seat per month with a two-seat minimum. For embedding conversational AI in a customer-facing product, use the OpenAI API instead, since consumer plans are not licensed for embedding. For regulated industries, only API with ZDR or a controlled deployment passes most audits.