ChatGPT Review: GPT-5.6, Work, Pricing and Best Uses

ChatGPT is becoming a stack, not a single chatbot

The most important change to ChatGPT in 2026 is not one benchmark result or one model name. It is the way the product is separating into several modes of work. Chat remains the fast conversational layer. Work handles longer tasks that may involve files, connected apps and multiple deliverables. Codex is the specialist coding environment. Voice has its own real-time model. Search and deep research handle live information. Projects, memory, tasks, plugins and connected tools sit around those modes as continuity and execution layers.

That makes ChatGPT harder to describe with the old question, "Which model does it use?" The model still matters, but users increasingly choose a surface and an effort level as well. A quick explanation, a long research job, a repository change and a spoken conversation can all happen under the ChatGPT name while using different systems, limits and operating assumptions.

For buyers, this is both the strength and the complication. One subscription can replace several separate AI products for a generalist user, but the plan table now matters more because model access, Work, Codex, voice, apps, memory and usage allowances do not expand at the same rate.

The August model change matters most at the lower tiers

OpenAI's August 6 update changed the everyday ChatGPT experience more than the July GPT-5.6 launch did for many users. GPT-5.6 Luna is now the default for Free and Go users, and both tiers receive unlimited everyday text chats subject to abuse safeguards. A Think option gives those users a way to ask Luna to spend more effort on a harder question, while limits still apply to uploads, images, data analysis and other tools.

This is a meaningful shift in the value of the Free plan. Free access is no longer best understood as a small sample of the main chatbot. It is a usable everyday text service with restricted access around the more expensive tools. Go, listed at $8 per month in the United States, is therefore less about unlocking a different default chat model and more about receiving higher limits for files, images, data analysis, memory and related features.

That distinction matters when comparing plans. A user who mainly asks text questions may find the Free tier much more capable than an older review suggests. A user who works with PDFs, spreadsheets, images or repeated research tasks can still hit the limits that separate Free from Go and Plus.

Free and Go are also the tiers that can carry advertising in supported markets. Ads are separate from the assistant's answers, but their presence is part of the product experience and should be considered when comparing the lower-cost plans with Plus and above.

Sol is now a controllable reasoning tier

GPT-5.6 Sol remains the flagship reasoning option for eligible paid ChatGPT plans. In August, OpenAI updated the version used in Chat and added a reasoning slider so users can choose how much effort the model applies. Sol powers several reasoning settings on eligible paid accounts, while Sol Pro is reserved for the highest-capability Pro experience.

This is a more practical model choice than a long list of unrelated model names. The user is deciding how much time and compute a task deserves. A straightforward rewrite does not need the same reasoning budget as a difficult architecture review, source synthesis or debugging problem.

The tradeoff is that more effort is not automatically better. Longer reasoning can use more allowance, take more time and still produce an incorrect conclusion. The appropriate test is task completion, not whether the model appeared to think harder. For coding, check the tests and diff. For research, open the sources. For analysis, reproduce the calculations. For writing, compare the result with the brief rather than assuming the higher setting improved it.

There is also an important product boundary. The August Chat update does not mean every GPT-5.6 surface changed at the same time. OpenAI states that Work and Codex can use versions of the GPT-5.6 family with their own availability and release cadence. A model name shared across Chat, Work, Codex and the API should not be treated as proof that every surface behaves identically.

Work changes the unit of value

ChatGPT Work is the larger product shift. It is designed for jobs that continue beyond a single answer. A user can give Work a goal, let it gather information from approved files, apps and the web, follow a plan, create finished materials and pause for clarification or approval when necessary.

That changes what should be measured. A chatbot is often judged by answer quality. An agentic workspace should be judged by the quality of the completed job, the amount of supervision required, the number of corrections, the evidence trail and the risk created by the actions it can take.

A useful Work evaluation therefore starts with a bounded real task. Give it a research brief with named source standards, a spreadsheet that needs a defined analysis, or a document package that has an objective completion checklist. Record how often it asks the right question, whether it respects instructions, whether connected information is used correctly, and how much manual cleanup remains.

Work can create documents, spreadsheets, presentations, reports and other finished outputs. It can also use scheduled tasks to keep work moving. Those capabilities are valuable, but they make permission design more important. A tool that can read connected material or act across a workflow needs narrower access and clearer approval points than a tool used only for brainstorming.

Codex is part of the product, but its economics are separate

Codex is now a major reason some users move from Plus to Pro, especially when they run long or parallel coding jobs. Current ChatGPT pricing gives Free and Go limited Codex access, expands it on Plus and provides the highest task capacity on Pro. GPT-5.6 availability also differs by Codex plan, with Terra available to Free and Go and a broader Sol, Terra and Luna choice on eligible paid tiers.

The main buying mistake is to treat a ChatGPT subscription as unlimited developer compute. Agentic coding can consume allowance quickly, and flexible credits or usage-based mechanisms can apply when included capacity is exhausted. Business and Enterprise pricing can add another layer because Chat, Work, Codex and related agentic products may draw from shared credit structures.

Developers should compare the cost of an accepted change rather than the number of prompts. A productive session is one that produces a reviewable diff, passes tests, respects repository constraints and does not create hidden cleanup. A cheaper model or lower plan can be better value if the task is well scoped. A more expensive tier earns its price only when the extra capacity reduces waiting, context loss or manual intervention enough to matter.

The OpenAI API is also separate from a ChatGPT subscription. Paying for Plus or Pro does not create an included pool of API tokens. A product team building its own application should budget the API independently rather than using the consumer subscription price as a proxy.

Voice is now its own interaction model

GPT-Live changed ChatGPT Voice from a turn-by-turn imitation of a phone call into a more continuous interaction. The Live experience can listen and speak at the same time, handle interruptions and use supported context such as web search, memory, text and images. Paid plans use GPT-Live-1 while Free uses GPT-Live-1 mini, subject to current limits and rollout conditions.

That makes voice more useful for situations where hands and eyes are busy, such as walking through a problem, practising a language, reviewing a plan or coordinating work from the desktop app. Voice can also be used with Work and Codex on supported desktop experiences.

The limits are important. Live does not automatically expose every feature available in text chat, and video, screen sharing, connected apps or plugins can depend on the selected voice mode, device and rollout. A product page should not collapse "Voice" into one universal capability.

For users who need voice as a core workflow, the test should include interruption handling, factual accuracy, latency, accent and language quality, device switching and what happens when the task needs a visual or file. A natural conversation can feel more trustworthy than text even when the underlying answer still needs checking.

Pricing is simple at the bottom and more complex at the top

The headline consumer ladder is straightforward. Free costs nothing. Go is $8 per month in the United States, with localized pricing in some markets. Plus is $20 per month. Pro now has two monthly tiers, $100 and $200, with the same core Pro capabilities but different usage allowances. OpenAI describes those allowances as roughly five times and twenty times Plus respectively.

The practical difference between Plus and Pro is therefore capacity as much as intelligence. Plus already includes advanced GPT-5.6 reasoning, expanded research, projects, tasks, custom GPTs, Codex and Work. Pro adds Sol Pro and materially higher limits for people who use those features heavily.

Business uses a different model. Standard seats are currently $25 per user per month on monthly billing or $20 per user per month when billed annually, with a minimum of two standard seats. Business adds a managed workspace, administration and business data protections, while agentic usage can be extended through workspace credits. Enterprise remains contract priced and adds deeper governance, security, deployment and support options.

These are US list prices and public plan descriptions, not a promise that every user sees the same checkout amount. Taxes, local currency billing, regional availability, staged rollouts and workspace configuration can change the actual experience. The correct buying step is to verify the live pricing page and the plan picker before committing.

Canvas is no longer the right way to describe ChatGPT writing

Older ChatGPT reviews often list Canvas as a central writing and coding feature. That description is now stale. OpenAI removed Canvas from the current GPT-5.5 Instant and Thinking experience in May and moved writing and coding interactions directly into chat through writing blocks and code blocks, while legacy Canvas access was temporary.

This is a good example of why a broad AI workspace needs more frequent maintenance than a conventional software listing. The capability did not disappear in the simple sense. The product interaction changed. A tool directory that keeps the old feature name can be technically familiar and still mislead a new user about what they will see.

The same rule applies to model names, agent labels and plan limits. ChatGPT changes often enough that a useful page should describe stable jobs first and release-specific implementation second.

More capability creates more boundaries to check

ChatGPT can now touch more kinds of work, which makes source checking and permission checking more important rather than less. Search and deep research can cite sources, but a citation can still be misread or fail to support the exact claim beside it. Files can be analysed, but extracted tables or document structure can be misunderstood. Code can be written, but it can still introduce regressions. An agent can complete a workflow, but completion is not the same as authorization or correctness.

Connected apps add another layer. The right question is not only whether ChatGPT can reach a service. It is which account is connected, which actions are permitted, what data is returned, what the model can retain in the conversation and where human approval is required.

For individual accounts, data controls also differ from business workspaces. The consumer pricing page states that model-improvement use can be opted out of, while business offerings provide different default data treatment. Anyone handling confidential or regulated material should read the current privacy and workspace terms rather than inferring protection from the plan name.

How to evaluate ChatGPT in 2026

The best test is not a collection of clever prompts. Choose three real jobs that represent why you might pay for the product.

First, test a quick recurring task in ordinary Chat. Measure whether Luna or Sol gives you a reliable result without unnecessary setup. Second, test a source-heavy or file-heavy problem where deep research, uploads or data analysis must produce a checkable answer. Third, test a longer task in Work or Codex where the system must maintain instructions across several steps and deliver something that can be accepted or rejected against a checklist.

Track correction time, source quality, failed tool calls, usage pressure and how often you need to restate context. If voice matters, repeat one task in Live. If plugins or connected apps matter, test with the least-privileged account that can still complete the job.

This exposes ChatGPT's real advantage. It is rarely that one model wins every isolated comparison. The advantage is the ability to move between conversation, research, files, media, coding and delegated work without rebuilding the entire workflow each time.

Where ChatGPT fits now

ChatGPT remains the strongest default candidate for people who want one general AI workspace rather than a collection of specialist products. Free and Go now cover much more everyday text use than older plan comparisons suggest. Plus is the practical centre of the product for users who want advanced reasoning and the broader tool set. Pro is mainly a capacity and top-model decision for heavy users. Business and Enterprise are governance and organizational decisions as much as model upgrades.

The weakness is the same breadth that makes the product useful. Plan differences are increasingly complicated, features can roll out at different times, agentic usage can have its own economics, and a polished interface can make probabilistic output feel more settled than it is.

The right reason to choose ChatGPT is not that it can do everything. It is that enough of your real work can happen in one coherent environment with acceptable quality, cost and oversight. If one specialist task dominates your day, a narrower product may still be the better tool. If your work moves constantly between research, writing, analysis, files, code, images, voice and follow-up actions, ChatGPT's expanding workspace is the more important story than any single model release.

WhatAI Decision Box

ChatGPT is OpenAI's broad AI workspace for conversation, research, files, data, images, voice, coding and delegated work. In 2026, the main decision is no longer just which model it uses, but which combination of Chat, Work, Codex, research tools and plan limits fits the job.

ChatGPT is strongest when one person or team moves between several kinds of work. A project can begin with live research, continue through file and spreadsheet analysis, become a written or visual deliverable, and then move into Codex or ChatGPT Work without rebuilding the entire context in separate products.

Breadth does not remove uncertainty. ChatGPT can misread sources, make incorrect calculations, introduce code defects or take an agentic workflow in the wrong direction. Important outputs should be checked against primary evidence, tests, calculations, permissions and the original brief.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-08-26.

ChatGPT Community Discussions

Explore community discussions. Ask and answer questions on ChatGPT to grow and learn together.

ChatGPT Pros & Cons

How to Get Results with ChatGPT: Step-by-Step Workflow

1. Choose the job before the model

   Define the actual outcome, evidence standard, acceptable error rate, privacy boundary and completion criteria before deciding whether the task belongs in Chat, deep research, Work or Codex.

2. Select the lightest suitable mode

   Use ordinary Chat for quick questions and drafting, deep research for evidence-heavy work, Work for multi-step deliverables and Codex for repository or software tasks.

3. Set reasoning effort deliberately

   Use higher GPT-5.6 Sol effort only when the task benefits from deeper reasoning. Keep routine rewriting, extraction and formatting on lower-cost or faster settings.

4. Provide authoritative context

   Upload the relevant files, name the sources that control the answer and state which material is background only. Do not rely on the model to infer the evidence hierarchy.

5. Constrain connected access

   Connect only the apps and accounts required for the job. Use the least privilege that still allows completion and identify actions that require explicit approval.

6. Ask for an inspectable output

   Require citations, calculations, assumptions, file references, test results or a change summary depending on the task so that the result can be checked.

7. Validate the risky layer

   Open important sources, reproduce key calculations, run generated code and inspect any external action before accepting the result.

8. Measure correction cost

   Track how much time is spent fixing facts, prompts, formatting, code or agent actions. Compare total accepted-work time rather than raw generation speed.

9. Tune plan capacity from real use

   Only move from Free or Go to Plus, or from Plus to Pro, after identifying which limits interrupt a recurring valuable workflow.

10. Recheck changing features

    Verify plan limits, model availability, app access and pricing before documenting or automating a workflow because ChatGPT changes frequently.

ChatGPT Gotchas and Limits to Know Before You Start

• GPT-5.6 availability differs between Chat, Work, Codex and the OpenAI API.
• Free and Go use GPT-5.6 Luna for everyday chat and do not include GPT-5.6 Sol.
• Unlimited everyday text chat remains subject to abuse safeguards and does not make uploads, images or other tools unlimited.
• Free and Go can show ads in supported markets, while paid higher tiers are ad-free under current plan policy.
• The $100 and $200 Pro tiers share core capabilities but have materially different usage allowances.
• ChatGPT Business requires at least two standard seats under current self-serve billing.
• Business agentic usage can involve workspace credits beyond included allowances.
• OpenAI API usage is separate from ChatGPT subscriptions and must be budgeted independently.
• Search citations make evidence easier to inspect but do not guarantee that the source was interpreted correctly.
• Deep research can retrieve weak or incomplete sources unless the user specifies evidence standards.
• Generated code must be tested and reviewed before deployment.
• Work and connected apps can create permission and approval risks that do not exist in ordinary chat.
• Voice feature support varies between Live, Advanced and Standard modes and across devices.
• Canvas is a legacy description for current writing workflows and should not be listed as a primary current feature.
• Model, feature and usage limits can change during staged rollouts, so the live plan page should be checked before purchase.

Which ChatGPT Feature Fits Your Use Case

How Well ChatGPT Fits Common Use Cases

Starter Prompts for ChatGPT

What is ChatGPT best used for in 2026?

ChatGPT is best used as a general-purpose AI workspace for research, writing, file analysis, data work, coding, images, voice and multi-step tasks. Its strongest advantage is moving between those jobs without changing products.

Which model does ChatGPT use on Free and Go?

GPT-5.6 Luna is the current default for everyday text chats on Free and Go. Both tiers can also use Think for harder questions, while GPT-5.6 Sol is reserved for eligible paid plans.

What is GPT-5.6 Sol?

GPT-5.6 Sol is OpenAI's flagship reasoning model for complex work across areas such as coding, research, knowledge work, science, computer use and design. Eligible paid ChatGPT users can select different reasoning effort levels.

What is ChatGPT Work?

ChatGPT Work is an agentic mode for longer tasks. It can research, use approved files and connected apps, carry out multiple steps and create finished outputs such as documents, spreadsheets, presentations and reports while allowing the user to review or redirect the work.

How much does ChatGPT cost?

Current US list pricing is Free at $0, Go at $8 per month, Plus at $20 per month and Pro at $100 or $200 per month. Business standard seats are $25 monthly or $20 per user per month on annual billing with a two-seat minimum. Enterprise pricing is custom.

What is the difference between the two ChatGPT Pro tiers?

The two Pro tiers have the same core Pro feature set. The main difference is usage allowance: the $100 tier provides about five times Plus usage, while the $200 tier provides about twenty times Plus usage under current plan rules.

Does ChatGPT Plus include OpenAI API usage?

No. ChatGPT subscriptions and OpenAI API usage are billed separately. A Plus, Pro or Business subscription should not be treated as prepaid API credit.

Can ChatGPT search the web and cite sources?

Yes. ChatGPT Search can retrieve current web information and provide source links, while deep research can handle broader multi-source work. Citations still need inspection because a source can be misread or only partially support a claim.

Can ChatGPT analyse PDFs and spreadsheets?

Yes. File uploads, vision and data analysis support documents, presentations, spreadsheets, images and other supported formats. Limits vary by plan, and important extracted figures should be checked against the original file.

Does ChatGPT still use Canvas?

Canvas is no longer the current default writing and coding surface. OpenAI moved current writing and coding workflows into chat through writing blocks and code blocks, while legacy Canvas availability was temporary during model retirement.

Can ChatGPT use voice?

Yes. The current Live voice experience is powered by GPT-Live on supported plans and devices. Voice capabilities, limits, video or screen-sharing support and access to other tools can vary by mode, plan and platform.

Is ChatGPT always accurate?

No. ChatGPT can make factual, logical, coding, calculation and source-interpretation errors. For consequential work, verify sources, test code, reproduce calculations and review agent actions before relying on the result.

Related AI Models: LLMs, Multimodal Systems, and More Tools

Explore the Network