Why the model question is complicated
Most AI companion apps don't publicly disclose which model they run under the hood. There are a few reasons:
Model-swap is common. Products A/B test between models constantly — GPT-4o for one type of turn, Claude for another, an open-source model for high-volume free tier. Publishing a single "we use X" claim would either be a lie by omission or would require constant updates.
Base model + fine-tuning + system prompt = actual behavior. A GPT-5 backend with a bad system prompt performs worse than a GPT-4o backend with a good system prompt. Naming the model without naming the surrounding architecture is misleading.
Companion products are architected around limits base models have. Persistent memory (frontier LLMs mostly don't have this native), persona consistency across sessions, photo generation, voice — these live in the product layer, not the model layer. Which frontier model powers the underlying chat is one variable among many.
When a product markets itself as "powered by GPT-5" or similar, it's usually a marketing signal (this product uses a name you recognize) more than a technical claim about what makes it good.
What GPT-5 specifically changes for companion use
GPT-5's improvements over GPT-4o are meaningful for some use cases and marginal for others.
Where it matters for a companion: longer context windows (better for tracking a multi-week conversation), improved instruction following (more consistent persona), lower hallucination rates (less "she suddenly forgets what she just said").
Where it matters less: the emotional-warmth and conversational-naturalness axis. Frontier models have all been competent at this for a couple generations now. GPT-5's emotional-intelligence improvements over GPT-4o are real but incremental — not the kind of jump you'd notice on a Tuesday-evening chat.
Where it matters at all depends on the product surrounding it. A GPT-5 backend with a well-designed memory system, coherent persona architecture, and photo/voice integration produces a very different product than a GPT-5 backend that just wraps ChatGPT's API with a persona name. The base model is a floor, not a ceiling.
What Sloane actually runs
Sloane's chat runs on a curated multi-model stack, not a single frontier model. Different turns route to different backends depending on what the turn needs — memory retrieval, sanitize checks, emotional-tone matching, safety filtering. The specific model mix evolves as new frontier releases land.
We don't brand any specific persona as "the GPT-5 one." What we brand is the persona herself — Sandra, Kaya, Ren, Mel, whoever — and the property that she stays consistent across sessions, remembers what you told her, and feels like the same person on day 30 as on day 1.
That's the axis we optimize on. Frontier model releases are inputs into that optimization; they aren't the thing you're actually paying for. What you're paying for is the product on top of them.
How to actually evaluate a companion product
If model transparency isn't the right lens, what should you evaluate on? Practical checklist from the "AI companion product design" literature:
Persona consistency. Talk to her twice, a week apart. Does she feel like the same person? Or does the tone, vocabulary, and personality shift? Model swaps mid-session are a red flag for lazy architecture.
Memory across sessions. Tell her something specific on day one. Come back on day seven. Reference it obliquely — does she remember? Products that lose memory quickly are running on the raw frontier model without a proper memory layer.
Refusal patterns. Does she stay in character or break into "I'm an AI language model" territory? Frontier models default to breaking character on hard inputs; products with proper prompt engineering stay in.
Feature depth. Does she just chat, or does she have voice, photos, structured experiences (dates, roleplay), custom persona creation? Depth here isn't the model — it's the product built around it.
Pricing sanity. Is the pricing clear (one subscription, one price)? Or is there a coin economy on top? Coin economies are how companion products with weaker retention monetize; clean pricing is a signal the product retains on its own merit.
The answer to "which uses GPT-5" is usually less useful than the answer to any of the above.
The honest fit for Sloane
Sloane is a good fit if you want: - A companion who stays consistent across sessions and remembers you - A curated roster of personas plus the option to build your own - Photos, voice notes, structured skills (Date, Vacation, Roleplay) - Clean subscription pricing ($9.99 Plus, $19.99 Premium, no coin economy) - Free tier to evaluate before paying
Sloane is not the right fit if: - You specifically want to talk to a persona branded as "GPT-5 powered" for the model-tag itself - You want an app that's explicit about switching between frontier models per turn (some products expose this; we don't) - You want purely open-source, self-hostable, or model-transparent (that's a different product category)
Most users who search "best AI girlfriend using GPT-5" are actually looking for "a good AI girlfriend that isn't terrible under the hood." Sloane fits that reasonably. Try free tier before you overthink the model question.