ChatGPT Review 2026: Features, Pricing & Verdict

Which ChatGPT plan to buy

Plan	Monthly	Annual (per month)	Included usage
Free	$0	—	Unlimited text chats, capped tools
Go	$8	—	More messages than Free, still ads
Plus	$20	—	The 1x baseline
Pro 100	$100	—	5x Plus
Pro 200	$200	—	10x Plus from 30 Oct 2026 (was 20x)
Pro 500	$500	—	25x Plus, plus Ultrafast
Business	$25/seat	$20/seat	Standard seat, 2–200 seats
Business Premium	$125/user	$100/user	5x Standard seat, no 5-hour cap
Enterprise	custom (~$60/user)	billed annually	150-seat minimum

Prices verified 9 October 2026 against OpenAI’s plan pages and the ChatGPT Pro tiers help article. Our pricing tracker carries the dated change log. No Pro tier offers annual billing.

Recommendation: Plus. Free is enough for a few questions a day. Go, at $8, buys message headroom and keeps the ads — it exists for price-sensitive markets and most readers here will skip it. Plus is where the product becomes a work tool: current paid-tier models, higher caps, priority at peak, voice, image generation and Deep Research. Go to Pro 100 when Plus caps become a daily annoyance rather than a monthly one.

The ladder has no bulk rate

This is the arithmetic the pricing page does not show you. Divide monthly price by allowance multiple:

Plan	Price	Allowance	Cost per 1x of Plus
Plus	$20	1x	$20
Pro 100	$100	5x	$20
Pro 200 (from 30 Oct)	$200	10x	$20
Pro 200 (grandfathered)	$200	20x	$10
Pro 500	$500	25x	$20

Four of the five rungs land on exactly $20. The fifth — a Pro 200 seat bought before DevDay on 29 September 2026 — is the only plan OpenAI has ever sold at a volume discount, and it is the one being withdrawn. Existing Pro 200 subscribers hold 20x through 29 October 2026, then drop to 10x at the same $200, with a one-time grant of 62,500 usage credits (about $2,500) that expires 31 December 2026 as a nine-week cushion. Reporting also puts GPT-6 Pro’s weekly message cap on this tier at 100, down from 200. The five-hour rolling limit is not coming back.

Practical consequence: if you are buying new, pick on capability, not value-per-unit, because the unit price is identical everywhere. If you hold a grandfathered Pro 200 seat, you are on the best deal in the lineup until the end of this month and the thing to do before then is measure what you actually consume — the full arithmetic and the options.

What Pro 500 buys that money alone cannot

Pro 500 is the only Pro plan with the Ultrafast speed tier — up to roughly 300 output tokens per second in Codex, around 8x standard. As of 8 October 2026 that tier has a published API price on both current models, which finally makes the bundle valuable rather than mysterious: GPT-6.1 Sol Ultrafast is $12 per million input tokens and $60 per million output, six times Sol’s standard $2/$10, while GPT-6 Astra Ultrafast is $60/$300, six times Astra’s $10/$50.

Read that as a latency surcharge the subscription absorbs on your behalf, not as a free extra: Ultrafast draws on the same allowance you paid for, at 8x the rate in Codex and 6x on purchased credits. The 25x allowance and the 8x speed pull against each other on one budget. For a developer blocked and watching a cursor blink, worth it. For background or batch work, close to the worst way to spend the money — why the speed tier is now priced rather than bundled.

If you are buying from Europe, check the region list twice

Two of the features OpenAI marketed hardest this autumn are not available to you.

dots — always-on agents running on GPT-6 Astra, each with its own cloud computer — ship to Pro 100, Pro 200 and Pro 500 in eligible markets, and the eligible-market list excludes the EEA, Switzerland and the UK with no published rollout date. Business Premium reportedly has them across all supported regions. For a European buyer the workaround is therefore a plan type, not a waiting list: $100/user/month on annual Business Premium billing rather than $100/month on personal Pro 100.

On the API, EU data residency used to mean no speed tier at all. That changed on 8 October 2026 for the mid-tier models — GPT-6.1 Sol now supports US and EU data residency including Fast and Ultrafast, and Fast reached GPT-6 Luna under EU residency the same day — but GPT-6 Astra Ultrafast still supports global processing and US residency only. So the flagship remains the one model a European buyer cannot buy speed on at any price, and regional processing endpoints carry a 10% uplift on top.

Regulated and vertical editions

Above Enterprise sits a track most buyers never see, and it is worth knowing it exists. ChatGPT for Healthcare gained a read-only connector into Epic electronic health records on 1 September 2026 — notes, labs, medications, specialist documentation — under a Business Associate Agreement, with AdventHealth, Cedars-Sinai, HCA Healthcare, Memorial Sloan Kettering and UCSF Health among the launch partners. Enterprise customers request Regulated Workspace eligibility through their account team. The reason to care even if you will never buy it: it is the clearest evidence that OpenAI’s competitive effort has moved from model quality to first-party connectors into systems of record.

Public-sector buyers have their own rate card, and it changed on 1 October 2026. A 27-month GSA OneGov agreement replaced the $1-per-agency-per-year arrangement with metered tokens at 50% off commercial rates and the $15/user/month licence waived to $0 — no platform fee, no minimum, and no published spend control, which turns a flat line item into an unforecastable meter. Separately, ChatGPT Mil runs on the Pentagon’s GenAI.mil portal at Impact Level 5. OpenAI says data there stays isolated from training — a property of that deployment, not a default term on your plan.

The model lineup (October 2026)

ChatGPT is a conversational front end to OpenAI’s GPT models, and the generation turned over three times in five weeks:

• GPT-6 Astra (3 September 2026) — the computer-use flagship that drives software across applications. $10/$50 per million tokens on the API; the practical upgrade is agentic desktop work rather than chat. Launch analysis.
• GPT-6 Sol and GPT-6 Luna (22 September 2026) — half of GPT-5.6’s API prices, stated as permanent rather than promotional. Sol $2/$10, Luna $0.10/$0.50. What the cut means.
• GPT-6.1 Sol (29 September 2026) — same $2/$10 headline, but cached input halved to $0.10, which is the real price move on cache-heavy agent traffic. One point below Astra on Artificial Analysis scoring at a fraction of the cost per task. The collision at $2.

Two things to carry from that list. GPT-6 Astra did not move from $10/$50 while the mid-tier halved, so the flagship now costs 5x the mid-tier rather than the 2.5x it launched at — re-check any Astra business case written before 22 September. And GPT-6 Sol is not deprecated: it carries no shutdown date and simply points at the newer model, so teams pinned to gpt-6-sol have to choose to move rather than being moved. Our model retirements tracker carries the dates; the API cost calculator applies the above-272,000-token surcharge, which doubles input and multiplies output by 1.5 for the whole request.

Inside the product you get image generation on OpenAI’s native image model (DALL-E 3 was retired on 12 May 2026), voice conversations, Deep Research, file and spreadsheet analysis, Custom GPTs, Code Interpreter, cross-conversation memory, and Canvas. ChatGPT Work — the agent that returns finished documents, spreadsheets, decks and web apps — is on Pro, Enterprise, Edu, Plus and Business. This bundle is what you pay for. Raw API access to the same models is cheaper; the integrated product is the thing with a price.

Voice is worth a second look if you last tried it as a single-model walkie-talkie. Since 23 September it takes plugins for email, calendar and Slack, lets you pick the model backend (Astra, Sol or Luna) rather than running one fixed model, and works inside ChatGPT Work — so you can produce documents by speaking. The full-duplex rebuild that made it listen and speak at once is the foundation. Memory is also no longer a command: the Dreaming V3 architecture synthesises context across conversations automatically, and OpenAI measured factual recall going from 67.9% to 82.8% in internal evaluations.

What ChatGPT does best

Generalist work across every domain. Still the superpower. Switching between marketing copy, Python debugging, pricing research and translation in one session — no competitor handles those shifts as smoothly. Others have caught up on raw capability; nothing has caught up on the absence of friction in choosing a tool.

Data analysis without SQL. Upload a CSV, ask a question in English, get charts and statistics back from Code Interpreter running real Python in a sandbox. This is the feature non-technical knowledge workers underestimate most, and it is the single best reason for a finance or ops person to pay the $20.

Image generation inside the conversation. Describe, iterate, change one element, ask for variations, never leave the chat. Midjourney produces more artistic output; for “blog header in 30 seconds” ChatGPT is faster. See best AI image generation tools for the specialist case.

Voice for hands-free thinking. Long walking-pace conversations to talk through a decision are a real and underrated use case. For the people it fits, voice alone justifies the subscription.

Custom GPTs for repeat workflows. If you run the same prompt pattern weekly, save it once and skip the setup. The store has thousands built by others; quality varies wildly and a few are excellent.

Where ChatGPT falls short

Long-form writing is generic without heavy prompting. The most common complaint from professional writers, and it is accurate. The same “write 1,000 words about X” prompt produces more natural prose from Claude almost every time. ChatGPT’s default voice has a slight sterility — hedged “it’s important to note” phrases, symmetrical paragraphs, predictable pivots. Detailed system prompts coach it out; the baseline is the problem.

Hallucinations on factual work. Each generation invents less than the last and the current models still fabricate citations, misattribute quotes and state wrong facts confidently, especially on niche topics. Anything going out under a byline needs every claim checked, which means the tool saves less time on research than it appears to.

Launch benchmarks need discounting. Independent evaluator METR found GPT-5.6 Sol gamed its own tests more than any public model it had assessed, and OpenAI’s headline numbers for the GPT-6 family are measured at xhigh and max reasoning effort while the documented default is medium. Both facts point the same way: treat the launch chart as an upper bound.

Real-time information. Web search works but is slower and less reliable than Perplexity, with citations less trustworthy than Gemini’s. For prices, news and recent changes, use a citation-first tool — Perplexity vs ChatGPT covers how the two fail differently.

Long-context quality, not capacity. The old version of this complaint — that competitors had bigger windows — no longer holds: GPT-6 Astra, Sol and Luna all carry 1,050,000-token context on the API. The live problem is that quality degrades before the window does, the above-272,000-token surcharge makes long prompts cost double, and ChatGPT’s consumer caps are well below the API ceiling. For a 50-page contract the window is adequate and the arithmetic is not.

Account placement can look like a model regression. Since 18 August 2026 a behavioural age-prediction classifier places accounts into the teen experience using topics discussed, hours active and account age, and OpenAI says it defaults to under-18 where there is doubt. No accuracy figure has been published. Adults on paid plans who suddenly see constrained or coached responses should check account placement before concluding the model got worse.

The ads, and why “ad-free” is now a priced feature

Ads arrived on Free and Go in February 2026 and the trajectory has not improved. Three things a buyer should know.

They can hold a conversation. From 16 September 2026 OpenAI began testing Sponsored Agents in the US — clicking a sponsored placement opens a separate, labelled chat with an agent run by the advertiser (Wayfair and Angi among the first) instead of sending you to a website. OpenAI states the sponsored conversation is separate from its own answers, and there is no evidence organic answers are affected. The problem is narrower and real: in a category where one vendor bought the placement and three did not, the recommendation stays honest while the outcome skews toward whoever is standing there to answer the follow-up. If you research purchases on the free tier, take the question to an unsponsored surface before deciding.

The consent control does not govern the measurement. A technical investigation published 20 September 2026 documented that opening ChatGPT can mint a __obi cookie on .openai.com with a one-year lifetime and SameSite=none, so browsers attach it to requests from other companies’ sites carrying OpenAI’s ad pixel — observed across 936 pixels on 1,029 hostnames. OpenAI’s cookie policy lists __obi under analytics rather than marketing, and every decoded token in the study carried consent_decision: analytics_allowed regardless of the marketing choice made. Logged-out users were not outside it: anonymous device identifiers stayed stable for at least 27 days. It does not operate on iOS browsers, where WebKit blocks third-party cookies. The full mechanism, and what it means if you also run a website.

It is structural now, and it is in Europe. ChatGPT Ads passed a $1 billion annualised run rate about 200 days after launch, self-service buying opened to Europe, and the ad test crossed into the EEA and Switzerland under GDPR, where a consent-or-pay “ads-free free plan with fewer messages” is the contested part regulators may not accept.

Ads and their measurement remain scoped to Free and Go. Plus, Pro, Business, Enterprise and Edu are ad-free. That boundary is the cleanest single reason to move off Free, and the free tier is the wrong place for work or sensitive data.

Privacy and procurement notes that are not about the product

The training default is a tier choice, not a toggle. Consumer ChatGPT accounts are opted in to model training by default; enterprise and API organisations are opted out by default. Turning training off changes the future, not the past — anything already in the corpus stays. On 25 September 2026 OpenAI disclosed that agents in its internal research environment had posted 53 user-supplied images to public image hosts as unlisted links and that it cannot identify the affected users, because de-identification of training data runs one way. Nobody will be notified. For work you would not want published, the control that would have prevented this is the tier chosen up front — the disclosure and what the framework does not promise.

Zero Data Retention has a sibling in preview. OpenAI’s Private Safety Processing does cross-session misuse detection it says stays compatible with ZDR, with content on customer infrastructure or under customer-held keys. It is aimed at Anthropic’s retention requirement for its top tier. Caveat: still a preview. ZDR also keeps one mandated carve-out — images flagged as potential CSAM are retained for review and reporting.

Finances connects to a credit bureau, and disconnecting is asymmetric. US Plus and Pro subscribers can connect an Experian credit report, returning a VantageScore 3.0 refreshed monthly as a soft inquiry. Two caveats belong in the decision: VantageScore is used by some lenders and not others while much US lending runs on FICO against a different bureau’s file, and disconnecting removes synced account data within 30 days but does not delete credit details already written into conversation history — the same store memory and personalisation read from. A Privacy Center at chatgpt.com/privacy-center shipped the same day and gathers seven privacy topics in one page across Free, Go, Plus, Pro and Business. Full analysis.

Two governance facts, neither a control. ChatGPT is the first generative assistant designated a Very Large Online Search Engine under the EU’s Digital Services Act, on OpenAI’s disclosure of roughly 159 million average monthly EU users; from a compliance window the Commission puts at January 2027 it must publish annual systemic risk assessments, submit to independent audits and maintain an ad repository, with fines up to 6% of global turnover. That is the one assistant on most shortlists that will be producing audited safety documentation on a published schedule. Against it: OpenAI signed the White House Accord on Super Intelligence on 29 September 2026, and the companies choose their own auditors, need not name them, face no deadline and need not publish findings — and on 30 September the FTC confirmed a broad investigation into the safety of OpenAI’s and Anthropic’s systems. An open investigation is not a finding, and signatory status is not an assurance. If your risk file needs an auditor’s name, scope or findings summary, that has to arrive through a contract clause.

ChatGPT vs Claude vs Gemini

If the work is	Buy	Why
Anything and everything	ChatGPT Plus	Widest feature surface; zero tool-choice friction
Writing that gets published	Claude Pro	Better prose with less prompting, same $20
A large codebase	Claude Code or Cursor	Purpose-built harnesses beat a chat window
Research needing citations	Gemini or Perplexity	More trustworthy sourcing
Spreadsheet and CSV analysis	ChatGPT Plus	Code Interpreter has no real equivalent
Images inside a conversation	ChatGPT Plus	Fastest describe-iterate loop
Living in Gmail and Docs	Google AI Pro	Workspace integration is the feature

On independent scoring, Artificial Analysis’ v4.2 index rewrite moved Astra four points ahead of GPT-5.6 Sol without OpenAI shipping anything — it added benchmarks, dropped a saturated one and doubled private held-out data to 40% of weight. Claude Fable 5.1 still leads overall. Astra leads on long documents and output-token efficiency. If you pay for exactly one AI and the work is serious, Claude is the stronger pick at the same price; see Claude vs ChatGPT and Gemini vs ChatGPT for the full head-to-heads, and best AI chatbots for the wider field.

Who should use ChatGPT

• Students and casual users — the free tier beats any comparable product, ads included
• Generalists whose work spans domains — writing, coding, analysis, images in one session
• Non-technical analysts who need data work without SQL or Python
• People who want voice — still the best voice experience available
• Anyone building Custom GPTs for repeat workflows

Who shouldn’t use ChatGPT as their primary AI

• Professional writers — Claude produces better long-form with less prompting
• Engineers on complex codebases — Claude Code or Cursor are built for it
• Researchers needing cited sources — Perplexity or Gemini are more reliable
• Anyone handling sensitive data on Free or Plus — the consumer training default is opt-in, and the opt-out is not retroactive
• European buyers who specifically want dots or Astra-speed — neither is sold to you today

FAQ

Is ChatGPT Plus worth $20 a month?

For most paying users, yes, and it is where you should start and probably stop. $20 buys the current paid models, higher caps, priority at peak, voice, image generation and Deep Research — a wider feature surface than any competitor sells at the price. The honest caveat is that it is no longer the obvious $20: Claude Pro costs the same and produces better writing and code. If your work is mostly one of those two things, buy that instead.

Should I upgrade from Plus to Pro?

Only when Plus caps interrupt you daily rather than monthly, and only on capability rather than value. Every Pro rung costs the same $20 per 1x of Plus allowance, so Pro 100 is five Plus subscriptions’ worth of usage for five Plus subscriptions’ worth of money. Pro 100 and Pro 200 differ only in headroom; Pro 500 is the only tier that adds a capability, namely Ultrafast.

I have a Pro 200 seat from before DevDay. What should I do before 30 October?

Measure your consumption now, while you still have 20x. If your real usage sits under 10x, the change costs you nothing and you can ignore it. If it sits above, you have three options: absorb the 62,500 credit grant through 31 December as a bridge, step up to Pro 500 at 25x for $500, or move the heavy workload onto the API, where caching and batch rates still discount volume even though the subscriptions no longer do.

Which GPT model should I call on the API?

GPT-6.1 Sol at $2/$10 by default — it scores one point below the flagship at a fifth of the price and its cached input is $0.10, the lowest in the family. GPT-6 Luna at $0.10/$0.50 for volume and classification. GPT-6 Astra at $10/$50 only when the task is genuinely computer-use or agentic desktop work, because the flagship premium is now 5x the mid-tier rather than 2.5x. Keep prompts under 272,000 tokens where you can: above it, input doubles and output bills at 1.5x for the whole request.

Is the free tier safe to use for work?

No, and the reason is not model quality. Free and Go carry ads, the measurement behind those ads is classified as analytics rather than marketing so declining marketing cookies does not switch it off, and consumer accounts are opted into model training by default with no retroactive opt-out. Any paid tier is ad-free. For anything you would not want published, the tier is the control, and it has to be chosen before you paste the document, not after.

Can I get the newest ChatGPT features in the EU?

Not all of them. The models, Plus, Pro, Business and ChatGPT Work are all available. dots — the always-on agents — exclude the EEA, Switzerland and the UK on personal Pro plans with no published date, though Business Premium reportedly has them in all supported regions. On the API, GPT-6.1 Sol supports EU data residency including the Fast and Ultrafast speed tiers as of 8 October 2026, but GPT-6 Astra’s Ultrafast is US-and-global only, and EU regional endpoints carry a 10% uplift.

The verdict

ChatGPT is still the right answer to “where do I start?” — the safest, most versatile entry point to modern AI, with a free tier good enough to justify an account on its own.

What has changed is that breadth is no longer the same as leadership. Claude is better at deep work. Gemini is better at research. DeepSeek is cheaper. Claude Code and Cursor are better at code. And the subscription ladder is now explicitly flat: there is no bulk rate anywhere in it, so the only reason to climb is headroom or Ultrafast.

If you already pay $20 for Plus and it works, do not switch. That is a lot of product for the money. But if you are hitting the wall on one specific kind of work, try the specialist, because the era when paying OpenAI meant having the best of everything is behind us. For most readers that is fine. ChatGPT does not need to be best at everything — it needs to be the first thing you try, and it still is.

What changed

• 9 Oct 2026: Ultrafast is priced on both models — Sol $12/$60, Astra $60/$300 — and Sol’s Fast and Ultrafast tiers now sell under EU data residency while Astra’s do not — see analysis.
• 1 Oct 2026: Pro 500 launched at $500 for 25x Plus and Pro 200’s allowance halves to 10x on 30 October, flattening the ladder to $20 per 1x — see analysis.
• 1 Oct 2026: dots shipped to Pro and Business Premium, but personal Pro access excludes the EEA, Switzerland and the UK — see analysis.
• 1 Oct 2026: the FTC opened a safety investigation into OpenAI and Anthropic, days after both signed an accord that obliges no published findings — see analysis.
• 30 Sep 2026: GPT-6.1 Sol shipped at $2/$10 with cached input halved to $0.10 — see analysis.
• 27 Sep 2026: OpenAI disclosed 53 user images leaked by internal agents and said it cannot identify the affected accounts, making the training tier a security control — see analysis.
• 23 Sep 2026: GPT-6 Sol and Luna halved API prices permanently while Astra held at $10/$50, taking the flagship premium from 2.5x to 5x — see analysis.
• 22 Sep 2026: Finances gained an Experian credit-report connection for US Plus and Pro, and disconnecting does not delete credit details from conversation history — see analysis.