How to choose an AI companion in 2026: evidence, tests and costs

Research updated 26 October 2026. Documentation-based editorial guidance; these are proposed reader tests, not results of our paid hands-on testing.

First choose an activity, not a character picture

An AI companion may be a long-running conversational partner, a character you direct through fictional scenes, an image generator, or a voice interface. The sales pages can make those seem interchangeable. They are not. Decide which one would make the subscription worthwhile if every other feature disappeared. Someone who wants consistent dialogue should prioritise memory controls and chat access; someone who wants images must inspect token pricing, not just the advertised monthly fee.

A useful shortlist has one essential feature, one optional feature and one spending limit. Example: “I need continuity for a fictional detective story, might use occasional voice, and will not spend more than £20 in any billing month.” This excludes services where attractive media functions require an unpredictable number of top-up purchases.

Choose by what the documentation actually proves

Your main goal	Evidence to inspect first	What can go wrong
Long-running conversations	Memory controls, resets, exports, plan-specific context	A remembered fact exists but is not recalled when needed
Hands-free conversation	Supported spoken language, subscriber requirement, voice allowance	English-first output or chargeable audio credits
Generated pictures or clips	Per-action token rate, monthly allowance, regeneration cost	“Unlimited chat” mistaken for unlimited media
Build your own character	Backstory controls, memory editing, example messages	A complicated setup produces inconsistent dialogue
Predictable monthly cost	Renewal total, recurring credits, payment route	Introductory price hides a higher renewal

Specific differences matter. Kindroid documents persistent and retrievable memories and paid tiers with larger context. Nectar lets users create custom memories of up to 200 characters and star up to ten. Nomi describes short-, medium- and long-term memory but warns that retrieval is imperfect. None of these descriptions independently proves which model remembers best in a real-world contest.

A 30-minute evaluation you can repeat

Minutes 0–5: basic suitability. Check whether the service allows the desired type of conversation and whether the interface works on your device. Inspect the content policy before writing a character brief.
Minutes 5–12: continuity. Tell a fictional character three invented details: your fictional town is Alder Quay, your meeting is on Thursday, and the missing notebook is blue. Change the subject. Ask about the notebook later without repeating its colour.
Minutes 12–18: correction. Change the notebook from blue to green. Look for a way to edit any saved memory. A good outcome is not merely a correct answer once; it is fewer contradictions after correction.
Minutes 18–23: language. Ask for two short spoken or written responses in your intended language, with one idiom and one date. Text competence does not establish good speech recognition or pronunciation.
Minutes 23–30: checkout. Record the renewal amount, billing interval, taxes, whether cancellation ends access immediately, and whether an extra purchase is required for your chosen feature.

This is a buyer’s suggested test protocol, not results of testing these providers. Repeat it on different days if continuity is your primary reason to pay. Never use real health details, passwords or another person’s information as test data.

What the plans may conceal

OurDream’s July 2026 support article lists a Standard plan at US$19.99 monthly or US$119.88 charged annually, with a monthly Dreamcoin allocation. Its action guide lists voice calls at 50 Dreamcoins per minute. A user who wants 30 call minutes would spend 1,500 Dreamcoins before any image or video generation. The calculation shows why a monthly membership alone is not a useful price comparison.

Kindroid’s October 2026 documentation describes a three-day premium trial that requires payment details and converts unless cancelled; its free preview and trial have different model access. Candy AI’s published paid-user guide describes unlimited text chat plus 100 tokens each month, with various media and creation tools consuming tokens. These are examples of quite different commercial structures, not direct measures of product quality.

UK payment and privacy checkpoints

For a UK buyer, a dollar price on a provider’s help page is not a verified pound checkout price. Check the live total in GBP, exchange-rate treatment, VAT presentation where relevant, whether you are paying through an app store or the provider, and the cancellation route attached to that exact purchase. Keep a dated screenshot of the renewal screen. Avoid annual plans until the service has demonstrated real value over multiple sessions.

Treat anything typed into the companion as information potentially stored by an external service. Read its privacy information for retention, training uses, account deletion, third-party processors and international transfers. Prefer fabricated names when evaluating a service. UK data-protection rights can include access and erasure in relevant circumstances, but that does not guarantee an immediate removal of every backup copy.

A decision rule rather than a universal winner

If consistency matters most, start by examining Kindroid, Nectar and Nomi memory documentation. If images matter most, begin with the published cost structures of OurDream, Candy AI or Lovescape and build a monthly budget. If hands-free use matters, read the voice rules for Nomi or Kindroid and test English-language performance before trusting a marketing demonstration. Our reviews are documentation-based, not hands-on performance rankings; any unverified language, price or privacy claim remains unverified.