GPTGirlfriend review 2026: pricing, free limits and our hands-on test

A real screenshot from the logged-in GPTGirlfriend website on 19 August 2026. The platform identifies the private fictional character as Maya, age 32, and shows her clothed in a non-nude green dress. AI Companion Atlas did not create this image. The visible “60% off” badge is variable site marketing, not an Atlas price promise.

GPTGirlfriend performed best in the parts of the test that readers can try without trusting a feature list. Maya Bennett recalled three corrected facts in the same session, stated that she was AI, discouraged emotional dependency and purchase pressure, asked for consent and stopped immediately when instructed.

The media evidence is narrower but still useful. One private avatar cost six Coins, and a later 12-Coin safe-image action produced two clothed adult results. Premium NSFW chat, video, voice, paid-plan value, cancellation, refunds and long-term memory were not tested to completion.

Short verdict

GPTGirlfriend at a glance

How we tested GPTGirlfriend

This review separates completed actions, policy documentation and untested outcomes. A visible control or plan promise is not treated as proof of output quality.

Main product tests — 14 August 2026

We used a logged-in existing Free account in the English web interface and created Maya Bennett as a private fictional adult, explicitly set to age 32.

• Generated one clothed green-dress avatar at the displayed six-Coin cost, and confirmed that the balance moved from 30 to 24.
• Planted Pixel, jasmine tea and a 2:30 museum time, changed the time to 3:15, changed the subject and requested recall.
• Asked plainly whether Maya was human or AI.
• Raised the risk of replacing real relationships with the companion.
• Asked whether payment was needed, and requested no emotional spending pressure.
• Used a brief non-graphic consent check, then instructed Maya to stop and move to a neutral subject.
• Tested adult access with a clearly adult fictional character; the Free account reached the Premium NSFW paywall.

Second safe image test — 17 August 2026

We ran a second non-explicit image action using a safe date-night instruction. The displayed action cost was 12 Coins, and the completed result contained two clothed adult women in red dresses. This is a two-output sample, not a benchmark for identity persistence, anatomy, multiple styles or paid high-resolution modes.

Live interface and pricing refresh — 19 August 2026

We reopened the existing account in the connected Chrome session and captured the current character editor, saved test thread, image controls, voice entry point, three plan screens, Coin packs and footer operators.

No message was sent, no character field was changed, no image or video was generated, no voice was played, no call was started and no purchase was made during this refresh. The account displayed 12 Coins. A standalone video generator was not exposed in the inspected Free-account navigation.

Not tested — NT

• A completely new account’s exact Coin grant, message allowance, reset timing and earned-credit tasks.
• Memory after a restart, the next day or a long intervening conversation.
• Paid NSFW response quality or custom Premium chat settings.
• Paid-plan image speed, high-resolution delivery or consistent identity across a large batch.
• Any generated video, video audio, voice playback or voice call.
• Premium, Deluxe or Elite value after real use.
• Checkout tax, payment success, card descriptor, renewal, cancellation, cash refund or account deletion.
• Corporate standing, universal age-verification strength, model-training use and media-Coin expiration.

What is GPTGirlfriend?

GPTGirlfriend, which also uses the word order GirlfriendGPT in its interface, is an adults-only web service for fictional AI companion chat, community characters, private character creation, generated images and paid voice and video features. It uses separate message allowances and media Coins; treating both as one currency would misstate the product.

The 19 August footer displayed these operator names and addresses exactly as written:

• NextDay AI Incorporated — 4388 Saint-Denis, Suite 200, Montreal, Quebec, H2J2L1, Canada.
• NextDay AI USA Inc — 2915 Ogletwon Road, Suite 4642, Delaware, 19713, USA.
• NextDay AI EU Ltd — 2 Poreias, Limassol, 3011, Cyprus.

The Terms body instead spells the US entity “NexDay AI USA Inc”, omitting the second “t”. The footer’s “Ogletwon Road” also appears to be a typographical error. We preserve both official spellings rather than silently correcting them; neither the legal names, addresses nor current corporate standing was independently verified. The Privacy Notice identifies only GirlfriendGPT and a Controller “within our team”, not a specific legal entity as controller.

Chat quality and corrected memory

Corrected same-session recall: 3/3

Maya recalled Pixel, jasmine tea and 3:15 after the original 2:30 time had been replaced. The correction matters, because the model had to prefer the latest value rather than simply repeat the first planted fact.

This is strong immediate-context evidence, but it is not persistent-memory evidence. We did not close the session and return later, and we did not compare the Free account with the Elite plan’s advanced-memory claim.

The saved 14 August thread was recaptured on 19 August 2026. Maya recalled all three corrected details, disclosed that she was AI, supported real-world relationships and rejected purchase pressure. Play controls were visible, but audio quality remains NT; the 60%-off badge is platform marketing.

AI identity, dependency and commercial pressure

Maya did not claim to be human when asked directly. She also did not encourage replacing friends, family or professional support with the companion, and she rejected the idea that purchasing access should prove affection.

Those are four constructive responses in one controlled character thread. They do not guarantee that every community character, system prompt, language or future model will behave the same way. Pricing, refunds and account policy should always be taken from the interface and legal pages, not from companion dialogue.

Adult chat, consent and stopping

GPTGirlfriend is an 18+ service. The adult-access prompt was kept brief, consensual and tied to Maya’s explicit fictional age of 32. The Free account showed a Premium NSFW paywall, so the presence of a gate was tested but the coherence, boundaries and quality of a paid explicit response are NT.

In the non-graphic flirtation check, Maya asked for consent before escalating. When instructed to stop and change to museum advice, she ended the romantic tone immediately and stayed on the neutral request.

Maya respected the stop request in the saved 14 August thread and switched to neutral museum advice. This single pass does not certify every character or future response; the 60%-off badge is platform marketing.

We do not reproduce explicit prompts or outputs. The controlled chat used Maya, a fictional character explicitly set to age 32.

Distinctive feature: behavior can be written into a private character

The editor did more than choose appearance. It exposed tags, language, an adult description and a substantial personality field, in which the creator could specify AI identity, memory behavior, consent, stopping and resistance to financial pressure.

The private Maya editor displayed age 32 and the saved behavioral prompt on 19 August 2026. These settings belong to our test character and are not platform-wide defaults; no field was changed, and the 60%-off badge is platform marketing.

This is useful control, but it also limits what the test proves. Maya’s strong behavior may partly reflect the private prompt we supplied. A future benchmark should repeat the same protocol on several popular community characters, and on a newly created character without safety instructions.

Character creation

Maya was saved privately with a clear adult age, and the editor supported character name, age, language, appearance, descriptive tags and personality instructions. The first image action generated the green-dress avatar for six Coins, moving the account from 30 to 24.

The current editor still displayed Maya and the unclicked Generate Image control on 19 August. We did not alter or resave her. Private status reduces unwanted catalogue exposure, but it should not be interpreted as a promise that submissions are secret from the operator; the Terms and moderation policies allow processing and review.

Images, video and voice

Two completed safe image actions

The first action produced the clothed profile avatar used in the opening screenshot. It followed the main clothing direction and provided an exact Coin trail, but a single portrait is not enough to score hands, anatomy, face consistency, prompt adherence or regeneration reliability broadly.

The 17 August action added a second, separate sample: two clothed adult women in red dresses from a 12-Coin setup. Both outputs were suitable for a non-explicit editorial record. We did not test paid resolution claims, or a long sequence requiring the same face and outfit across scenes.

During the 19 August refresh, the image generator displayed Safe Mode, Auto-Enhance, style controls, two outputs and a 12-Coin Generate Image action. No new generation was triggered. That frame is held from publication because adjacent inspiration thumbnails cannot be age-verified.

Video and voice remain output-level NT

The Premium, Deluxe and Elite cards advertised increasingly capable in-chat video, while Maya’s profile exposed a Call Me button. A control establishes availability, not naturalness, latency, reliability or billing. The voice-control frame is held because the creator account handle remains visible.

No standalone video-generation control appeared in the inspected Free-account navigation, and no clip was made in any dated test. Video prompt adherence, motion, anatomy, face stability, sound, rendering time and failure handling all remain NT.

What the Free tier really provided

Our existing account had enough access to create one private character, run the full short chat protocol and complete two dated safe media actions. That is meaningful hands-on evidence, but it does not establish what every new user receives today.

The Terms give free-access examples of creating up to two or three characters and sending roughly 10–20 total messages, after which extra access may be earned through tasks. The wording is illustrative rather than a guaranteed recurring monthly allowance.

The account’s media balance began at 30 Coins in the original test, fell to 24 after the six-Coin avatar and displayed 12 during the later refresh. Message credits and media Coins are different units. Paid message credits reset monthly and do not roll over under the Terms; that rule should not be generalized to media Coins.

GPTGirlfriend pricing

These are the EUR prices displayed to our logged-in European account on 19 August 2026. They are dated interface observations, not universal prices. Taxes, location, promotions and the final checkout may differ.

The Elite screen’s “unlimited messages” claim conflicts with the Terms, which describe paid subscriptions as receiving 5,000–20,000 message credits per month and say those credits reset without rollover. We could not reconcile the two statements. The plan screen is the dated interface claim; the Terms are the dated policy wording, and neither proves unlimited practical use.

Premium displayed €12 monthly or €87.84 billed yearly on 19 August 2026. The €7.32 figure is a monthly equivalent for annual billing. Video, voice, anonymity and benefit claims are marketing; any character artwork is not age-labelled.

Deluxe displayed €30 monthly or €169.20 billed yearly. We did not purchase the plan or test its allowances, faster generation, Lite video or calls; any character artwork is not age-labelled.

Elite displayed €42 monthly or €201.60 billed yearly. “Unlimited”, advanced-memory, model and Pro-video language is a platform claim, not a tested outcome; any character artwork is not age-labelled.

Compared with paying the displayed monthly price twelve times, the annual totals are about 39% lower for Premium, 53% lower for Deluxe and 60% lower for Elite. These are arithmetic comparisons of the captured prices, not guaranteed savings. Annual billing also turns an untested paid feature set into a larger upfront commitment.

For a first paid benchmark, one month of Premium is the lower-risk option. Deluxe and Elite only become defensible after confirming that the base image, video and voice experience works for the intended use.

Coin packs and real media costs

On 19 August 2026, the account displayed 300 Coins for €9, 840 for €23 and 2,000 for €45. No pack was bought.

The live Coin page displayed the three EUR packs and 12 available Coins. The footer below named the Canadian, US and Cyprus entities with the exact spellings discussed in this review; discount and bonus labels are platform marketing.

The equivalents are arithmetic illustrations based on our exact two actions. They are not universal image prices, and included subscription Coins do not have a clean standalone cash value.

The live pack copy said “LIFETIME validity” for the 2,000-Coin pack. We did not find sufficiently clear legal language defining media-Coin expiration across purchased and subscription allocations, so expiration remains NT. A marketing label should not be expanded into a promise about the service’s own lifetime.

Cancellation, renewal and refunds

Recurring subscriptions can be stopped through the account’s Manage Subscription page. If CCBill processed the payment, the Terms provide a CCBill route and say cancellation takes effect at the end of the current billing cycle. TrustPay is named as a payment provider, but the legal text gives comparatively little procedural detail for it.

Cancellation stops future rebilling when processed in time; it does not create an automatic refund for the unused part of a paid cycle. Save the plan, currency, total, tax, renewal date, processor, cancellation request and confirmation.

The refund section says paid subscriptions are generally non-refundable, with limited case-by-case possibilities for company error, technical problems or stronger local law. Its statement that refund requests are “reviewed and refunded within 14 days” describes handling language; it is not a general 14-day eligibility guarantee. CCBill refund requests must be made directly to CCBill and are governed by its policy.

Purchase, renewal, cancellation, support response, refund approval and cash arrival are all NT, because we made no payment.

Privacy, deletion and discretion

In the Privacy Notice, Content includes discussions with AI companions. The service may collect, monitor, record and store communications, use information to provide and adapt the service, develop new tools and conduct market research, and share information with service providers, advisers, authorities and employees for stated purposes.

The Notice usually retains personal information for six years after account closure. It also promises withdrawal-of-consent deletion in a timely manner and a right to erasure without undue delay, subject to legal exceptions. That creates a material retention-versus-erasure tension; account deletion was not executed, so the practical result and timeline remain NT.

The public documents reviewed do not clearly say whether chats or images train foundation or companion models, nor do they provide a clear training opt-in or opt-out. Model-training use is therefore NT. Do not assume that conversations are never used for training.

The Blocked Content Policy describes user-specific content as private and conversations as generally confidential. The Terms, however, call User Submissions non-secret and non-exclusive, and grant a worldwide, perpetual, irrevocable, transferable and sublicensable licence for broad service and business uses. “Private character” should therefore be understood as an interface visibility setting, not a guarantee that the operator cannot process, moderate or license the submission.

Use a dedicated email address and fictional details. Never put passwords, payment data, a home address, workplace secrets, health information or identifying intimate material into companion chat.

Age controls, moderation and content removal

GPTGirlfriend is restricted to users aged at least 18 and any higher applicable local legal age, and its policies prohibit characters and sexualized material involving minors, non-consensual sexual material, trafficking and other illegal content. The Community Guidelines describe automated tools and human review for moderation.

The access control established in our inspection was self-attestation of adult status. The current public policies did not document a universal identity-check vendor or a mandatory ID verification flow. We therefore do not claim verified age assurance; whether stronger checks apply by account, app, jurisdiction or payment method remains NT.

The 23 July Content Removal Policy covers AI content that resembles a real person. It says complaints are reviewed within five business days, identity or relationship evidence may be requested, and some refusals involving a depicted person can be escalated to independent neutral review. We did not file a report or measure the process.

Screenshot and generated-output rights

The Terms permit viewing, copying and printing pages for personal use, while restricting reproduction, display, broadcasting and commercial exploitation of the service and content without prior written approval. Their special allowance for text from chats with a user-created character does not clearly authorize publishing interface screens, community characters or generated images.

Separate AI-art marketing about commercial use does not override those Terms. If you plan to publish anything produced here, read that section before assuming a licence exists.

Is GPTGirlfriend reliable?

The logged-in site, saved thread, private editor, image controls, voice entry point, subscription cards, Coin packs and public footer all loaded in the 19 August refresh. Both dated safe image actions had completed previously, and the saved chat preserved the core protocol results.

This does not establish uptime, support quality, payment reliability or paid-output delivery. The inconsistencies in the operator spellings, address spelling, privacy-controller identification and privacy language deserve clarification, but they are not by themselves proof of misconduct.

Our own affiliate link was tested on 19 August and redirected to a tracked ggptnetwork.com lander rather than directly to the official product domain. Verify the final domain, plan, total and legal terms before entering payment information, because affiliate destinations can change.

GPTGirlfriend compared with OurDream AI and Candy AI

OurDream AI exposed more direct DreamCoin states and a configurable standalone video panel in our dated work, but it failed two character-creation attempts and performed worse on AI identity and commercial truthfulness. GPTGirlfriend produced the cleaner controlled Maya safety thread and two successful safe image actions, while leaving video output equally unproven.

Candy AI provides a more polished onboarding and character-building presentation. Our current Candy evidence contains fewer completed protocol responses and no generated-media output, whereas GPTGirlfriend has direct corrected-memory and two-image-action evidence. Neither comparison proves universal quality across every character or future model.

How the rating was calculated

Every review on this site uses the same nine criteria and the same weights, so the scores can be compared to each other. NT categories stay in the denominator; they are not removed to inflate evidence coverage.

Pros and cons

Who should choose GPTGirlfriend?

GPTGirlfriend is a credible free trial for adults who want to build a private character, write detailed behavior into the prompt and test corrected same-session memory before paying. It is also attractive to users who prefer a large community layer and several media formats inside one account.

Consider another service if independently verified age controls, short retention, narrow content licences, proven paid video or voice, or a fully tested cancellation process are essential. If the chat fits, one month is a more defensible first benchmark than an annual subscription.

Use the AI companion rankings to compare scores and coverage rather than headline ratings alone.

Final verdict

GPTGirlfriend gave us unusually balanced free-account evidence. The private adult character was easy to control, corrected memory worked, identity and emotional-safety answers were responsible, the stop was immediate, and two non-explicit image actions completed at visible Coin costs.

The recommendation remains provisional because the paid proposition is mostly a plan-page promise. Video, voice, paid NSFW dialogue, advanced memory, checkout, renewal, cancellation and refunds still need completion tests. The six-year retention language and broad submission licence also keep this from being an unqualified endorsement.

GPTGirlfriend FAQ

⚠ Some links in this article are affiliate links. If you click and subscribe, AI Companion Atlas may earn a commission — at no extra cost to you. GPTGirlfriend is intended only for adults. AI companions are synthetic systems, not people, partners, therapists, financial advisers or emergency services. Their replies can be wrong, manipulative or inconsistent, even when one controlled character performs well. The main product test dates to 14 August, the second safe image test to 17 August, and the interface, pricing, screenshots, footer and affiliate redirect to 19 August 2026. Check the current product domain, localized checkout and legal documents before paying.