OpenAI runs two separate products under one brand: ChatGPT, the consumer assistant sold as a monthly subscription, and the API platform, billed per token for developers. This page explains where the free tier ends, what each paid plan unlocks, and how the data policies differ.
Used this tool? Rate it
Used this tool? Rate it
OpenAI describes itself in its own terms of use as an AI research and deployment company whose mission is to ensure that artificial general intelligence benefits all of humanity. For anyone arriving at openai.com looking for a tool to use today, that mission statement is less useful than a practical fact: the company sells two quite different things, and confusing them is the single most common source of frustration for new users.
The first is ChatGPT, the consumer-facing assistant. You sign in, you type, you get answers. It is sold as a monthly subscription with a usable free tier, and it is what most people mean when they say they "use OpenAI".
The second is the API platform, aimed at developers and companies who want to embed the same underlying models into their own software. It has no chat window of its own. It is billed per token consumed, has its own dashboard, its own documentation site, and — critically — its own data policy.
These two are not connected by billing. A ChatGPT Plus subscription grants you nothing on the API. API credits do not raise your ChatGPT message limits. If you take away only one thing from this page, take that.
Around these two pillars sit several other products: Codex for agentic software engineering, ChatGPT Work for workplace agents, and business-tier workspaces. The product surface moves quickly, and some well-known names have already been retired — a point worth reading carefully in the sections below before you plan around any of them.
A useful mental model is to treat "OpenAI" as a research lab that ships two storefronts.
The consumer storefront optimises for immediacy. There is no capacity planning, no key management, no rate-limit engineering. The trade-off is that you accept the interface, the usage caps and the defaults that come with your tier.
The developer storefront optimises for control. You choose the model, the reasoning effort, the context length, the retention policy and the region. The trade-off is that you build everything around it yourself, and you pay for exactly what you consume rather than a flat monthly fee.
Neither is a discounted version of the other. They are priced on different axes, and they answer to different privacy defaults.
The current generation is the GPT-5.6 family, which is split into three named variants: Sol, Terra and Luna. This naming replaced the earlier numeric-suffix conventions and is worth learning, because plan documentation now refers to the names rather than to a single version number.
Independent coverage from technology press dated the family's public launch to 9 July 2026. Model naming in this space changes often, so treat any version number you read elsewhere — including in older reviews — as provisional until you check the current plan comparison table yourself.
The core of ChatGPT remains a text conversation with memory. It handles drafting, rewriting, summarising, translating, explaining and structured analysis. The practical strength is not any single task but the low cost of iteration: rejecting an answer and refining the request is fast enough that you can converge on something usable in a few turns.
Memory carries context between sessions on paid tiers, and a temporary-chat mode exists for conversations you would rather leave no trace of.
Not every question deserves the same amount of computation. The product distinguishes between fast "instant" responses and slower "reasoning" responses that work through a problem before answering.
This distinction has a direct effect on what you get. Instant replies suit lookup, rewriting and light drafting. Reasoning modes suit multi-step logic, mathematics, debugging and anything where a confidently wrong answer would cost you real time. Access to the stronger reasoning models is one of the main things that separates the paid tiers from the free one.
Context windows also scale with the plan: GPT Instant total context runs 27K on Free, 54K on Go and Plus, and 128K on Pro. Reasoning contexts are considerably larger, reaching 256K on the middle tiers and 400K at the top.
In everyday terms the published comparison translates these into approximate page counts — roughly twelve pages of text on the free tier's instant mode, around forty pages on the middle tiers, and around 250 pages at the top. If your work involves dropping long reports or entire codebases into a conversation, this row of the comparison table matters more to you than the model name does.
The platform is not text-only. Across tiers it covers image generation, vision (understanding images you upload), voice conversation, voice with video, file uploads and data analysis. The free tier includes limited and slower image generation; paid tiers raise both the quality ceiling and the throughput.
Beyond conversation, the product now includes scheduled tasks, projects for grouping related work, shared projects for collaboration, a built-in browser, connected apps that reach into external tools, and workspace agents on business tiers. Codex covers agentic software engineering and has its own usage allowance.
For developers the surface is different: a Responses API, agent SDKs, function calling, web and file search, structured outputs, fine-tuning, evaluations, batch processing and moderation endpoints. Requests can be tuned by reasoning effort, and cost controls include prompt caching, a batch tier and flexible processing modes.
The most common use remains prose work: first drafts, tightening long text, adjusting register for a different audience, and turning rough notes into something structured. It is most valuable where the cost of a blank page is high and the cost of checking the output is low.
Code generation, review, refactoring and debugging are heavily represented in real usage. Community reports after the GPT-5.6 launch described noticeably better one-shot builds compared with the previous generation, alongside cost efficiency gains on coding workloads. Codex extends this into autonomous multi-step engineering tasks.
Deep-research modes gather and organise material across sources. This suits literature scans, competitor overviews and background briefings. It does not remove the need to verify — it changes the job from finding material to checking it.
Uploading a spreadsheet and asking questions about it works well for exploratory analysis, chart generation and cleaning messy data. The context and upload limits of your tier are the practical constraint here.
On the API side, the dominant pattern is embedding the models into a product: support assistants, in-app copilots, classification pipelines, semantic search and content generation at volume. The per-token model makes unit economics calculable in advance, which is precisely why product teams choose it over a seat-based subscription.
Business and Enterprise workspaces connect to internal sources — document stores, ticketing systems, code repositories — so answers draw on company context rather than public knowledge alone, with administrative controls over who can enable what.
Start here, because it determines everything else.
If a person will be typing the questions, you want ChatGPT and you want a subscription. If software will be sending the requests, you want the API and you want per-token billing. If both are true — a team that also ships an AI feature — you need both, separately.
The temptation is to always use the strongest model. In practice, routing easy work to the cheap tier and reserving the frontier model for genuinely hard problems produces a better cost-to-quality ratio than running everything on either extreme.
The most reliable quality improvement is supplying the constraints, the audience and the format you want before the first attempt, rather than correcting across five turns afterwards. Examples of what "good" looks like outperform adjectives describing it.
Reasoning modes cost more and take longer. Use them where a wrong answer is expensive — logic, maths, security-adjacent code — and skip them where it is not.
Cached input is billed at a fraction of the standard rate, and batch processing costs meaningfully less than real-time calls. Workloads that repeat a long system prompt, or that tolerate delay, can cut costs substantially without changing model.
Treat output as a competent draft from someone who will not tell you when they are unsure. For dates, figures, citations, legal points and medical claims, verification is not optional.
If you use ChatGPT for anything sensitive, decide consciously whether your conversations should improve future models. The default differs from the API default, and the control is not in the same place as your billing settings.
The free tier serves casual use well. People who use it daily for writing or study generally find one of the lower paid tiers pays for itself in time saved.
Anyone doing sustained knowledge work — analysts, engineers, marketers, researchers — is the core paid audience. The mid and upper tiers exist for people whose bottleneck is model quality or throughput rather than access.
The API suits teams embedding language capability into their own software, where per-token economics and model control matter more than a polished chat interface.
Business tiers target organisations that need administrative control, single sign-on, centralised billing and contractual data commitments. Enterprise adds security controls, data residency and negotiated terms.
If you need guaranteed factual accuracy without review, offline operation, or a fixed cost per outcome rather than per token, this is a poor fit. Heavily regulated workloads may also require contractual arrangements beyond the self-service tiers.
ChatGPT is available on the web, iOS and Android across every consumer tier, and the iOS app is published by OpenAI OpCo, LLC with a 12+ content rating. The listing carries an average rating close to 4.8 across more than nine million ratings, and requires a reasonably recent iOS version.
A desktop application is also available, and it has become more central to the product recently: capabilities previously delivered through a standalone browser have been consolidated into the desktop app and a browser extension.
OpenAI confirmed in July 2026 that it would sunset the standalone ChatGPT Atlas browser and fold its agentic browsing capabilities into a Chrome extension and the ChatGPT desktop app. If you encounter guides recommending Atlas as a separate download, they predate this change.
The Sora web and app experiences were discontinued on April 26, 2026, and the Sora API is scheduled to shut down on September 24, 2026. Video generation content that once lived there is no longer available through that route, and any integration still calling the Sora API needs a migration plan before the September date.
For developers, access is through HTTPS endpoints with official SDKs, plus availability through certain cloud marketplaces where billing runs through the cloud provider and rates may differ from direct pricing.
Two separate billing systems operate here. Read both sections.
ChatGPT consumer pricing runs from Free at $0, to Go at $8 per month, Plus at $20 per month, and Pro starting from $100 per month.
Free provides unlimited everyday text chats subject to abuse guardrails, with limited messages and uploads, limited and slower image generation, limited voice chats, limited deep research, limited memory and context, and limited access to Codex and the desktop workplace features. The default model at this tier is the fast, low-cost variant.
Go adds more messages with tools, more uploads, more image creation, more voice chats and longer memory. Published plan notes indicate this tier may include advertising.
Plus is the first tier that unlocks the advanced reasoning models, along with expanded messages and uploads, more capable image creation, expanded deep research, larger memory and context, projects, scheduled tasks, custom GPTs, expanded Codex usage, broader access to the workplace app across desktop, web and mobile, and early access to new features.
Pro is described as offering either 5x or 20x more usage — the reason it is listed as a starting price rather than a single figure — plus the professional reasoning variant, maximum Codex tasks, unlimited and faster image creation, maximum deep research, and the largest memory and context.
ChatGPT Business is listed at $20 per user per month billed annually, or $25 per user per month billed monthly, and requires at least two users. It includes access across desktop and mobile, connections to tools such as Microsoft 365, Google Drive, Slack, GitHub, Linear and Figma, custom team agent plugins, centralised billing, usage analytics and spend controls, SAML single sign-on with MFA, and no training on business data by default.
Enterprise is custom priced and adds an expanded context window, SCIM, enterprise key management, domain verification, role-based access control, custom data retention policies, data residency support across ten regions, 24/7 priority support with SLAs, custom legal terms and volume discounts.
API pricing is entirely separate and consumption-based. On the standard short-context tier, GPT-5.6 Sol is listed at $5.00 per 1M input tokens and $30.00 per 1M output tokens, GPT-5.6 Terra at $2.00 and $12.00, and GPT-5.6 Luna at $0.20 and $1.20.
Several modifiers apply. Cached input is billed far below the standard input rate, which rewards stable system prompts. Long-context requests are billed at higher rates than short-context ones. A batch tier prices asynchronous work below real-time, and data-residency endpoints carry a 10% uplift for eligible models released on or after 5 March 2026.
Because rates have been adjusted more than once during 2026, treat any figure quoted in a third-party article as potentially stale and confirm against the official pricing page before you build a budget on it.
Anthropic's Claude is the most direct competitor for both chat and API use, with a similar subscription-plus-API structure and a reputation among developers for long-document handling and coding.
Google Gemini competes on multimodal breadth and integration with Google's productivity suite, which can be decisive for organisations already standardised on Workspace.
Microsoft Copilot delivers closely related capability embedded inside Microsoft 365, which suits companies who want AI inside existing workflows rather than a separate destination.
Open-weight models — including families you can self-host — trade some frontier capability for control, on-premise deployment and freedom from per-token billing. For privacy-constrained or high-volume workloads this trade can be economically compelling.
Perplexity focuses specifically on search and cited answers rather than general assistance, and is a better fit when provenance matters more than open-ended generation.
The honest summary is that the frontier labs are close enough that the deciding factors are usually integration, price at your volume, and data terms, rather than raw capability.
Reviewers frequently praise the low barrier to entry and conversational quality while flagging that answers can be outdated or confidently wrong. This is a structural property of the technology, not a bug awaiting a fix. The failure mode is dangerous precisely because the tone of a wrong answer is identical to that of a right one.
Community discussion after the GPT-5.6 release noted that workplace chat and Codex draw on a shared usage allowance, so ordinary desktop conversation can quietly consume the coding budget. Staged model rollouts have also meant that some subscribers saw a newly announced model appear in their account later than others.
Two prominent products have been withdrawn during 2026: the standalone Atlas browser and the Sora video experience, with the Sora API following in September. Anyone planning an integration should confirm a product's current status rather than relying on documentation or reviews written earlier in the year.
OpenAI requires users to be at least 13 years old, or the minimum age of consent in their country, and anyone under 18 needs permission from a parent or legal guardian.
Per-token billing scales with success. A feature that becomes popular becomes expensive, and without spend limits and caching strategy the bill can move faster than the product plan anticipated.
Availability, data residency options and contractual terms vary by region and plan. Organisations with strict compliance requirements should confirm that the tier they are buying actually carries the commitments they need, rather than assuming the brand-level commitments apply to the self-service plan.
No, and this is the most frequent misunderstanding. ChatGPT subscriptions and API usage are billed through entirely separate systems. Paying for Plus, Pro or Business gives you no API credit, and buying API credit does not lift your ChatGPT message limits. If you need both, you pay for both.
The free tier gives unlimited everyday text chats subject to abuse guardrails, with limits on messages with tools, uploads, image generation, voice chats, deep research, memory and context. It runs the fast low-cost model rather than the advanced reasoning models, which are reserved for Plus and above. For casual use it is genuinely sufficient; for sustained professional work the caps become the constraint.
The current family is GPT-5.6 in three variants. Free and the entry paid tier default to Luna, the fast low-cost variant. Plus unlocks Sol, the advanced reasoning model, and Pro adds the professional reasoning variant. A mid-tier Terra model sits between them. Because these assignments change with product updates, the plan comparison table is the authoritative answer at any given moment.
It depends entirely on which product you use. For individual services such as ChatGPT, OpenAI states that it may use your content to train its models, and that you can opt out through the privacy portal by selecting the option not to train on your content. By default OpenAI does not train on any inputs or outputs from its business products, including ChatGPT Team, ChatGPT Enterprise and the API Platform. Consumer defaults and business defaults are opposites, which is why this question cannot be answered with a single yes or no.
By default, abuse monitoring logs are generated for all API feature usage and retained for up to 30 days, and eligible customers can apply for Zero Data Retention or Modified Abuse Monitoring controls. Those controls require prior approval and acceptance of additional requirements, so they are not simply a settings toggle.
As between you and OpenAI, you retain your ownership rights in the input you provide and you own the output you receive. That said, ownership under the terms is not the same as a guarantee of originality or of freedom from third-party rights, and similar prompts can produce similar output for different users. Commercial use should still be reviewed against your own risk tolerance.
Pricing is per million tokens and differs by model and context length. On the standard short-context tier the flagship is $5.00 input and $30.00 output per 1M tokens, the mid model $2.00 and $12.00, and the low-cost model $0.20 and $1.20. Cached input costs a fraction of standard input, batch processing is cheaper than real-time, and long-context requests cost more. Because these rates were adjusted during 2026, verify against the official pricing page before budgeting.
Both were withdrawn during 2026. The Sora web and app experiences were discontinued on April 26, 2026, with the Sora API scheduled to shut down on September 24, 2026. The standalone Atlas browser was sunset in favour of the ChatGPT desktop app and a browser extension. If you are following a guide that treats either as a current product, that guide is out of date.
Users must be at least 13 years old, or the minimum age required in their country to consent, and anyone under 18 must have permission from a parent or legal guardian. The mobile app additionally carries a 12+ store content rating, which is a separate classification from the terms-of-use requirement.
For business use the relevant tiers are Business and Enterprise, which carry a default commitment not to train on business data, along with SAML single sign-on, administrative controls and the ability to execute a data processing addendum. Enterprise adds custom retention, data residency and negotiated terms. Consumer subscriptions do not carry the same defaults, so putting confidential material into a personal account is a different risk decision entirely.
