Which AI models answer, and why
Every AI assistant answers through a specific model, and Sigtrip measures your hotel on each assistant through one of them. Here is which model that is today, why we chose it, and what happens when it changes.
The models answering today
| Assistant | Model | Switched on | Why |
|---|---|---|---|
| ChatGPT | GPT-5.6 Luna | Jul 31, 2026 | GPT-5.6 Luna is the model ChatGPT gives free users by default, so it's closest to what most travelers see. |
| Perplexity | Sonar | Apr 30, 2026 | Sonar is the model behind Perplexity's default free search, and like the app it searches the web for every answer. |
| Google AI | Gemini 3.1 Flash-Lite | Jun 9, 2026 | Google AI's free app is moving to a newer Flash-Lite model; we keep 3.1 Flash-Lite until we've compared the two on real traveler questions, so your trend isn't broken by an untested switch. |
| Claude | Claude Haiku 5.5 | Oct 8, 2026 | Claude's free app uses a larger model (Sonnet 5); we answer with Haiku 5.5, a smaller Claude model that named as many hotels as Haiku 4.5 when we compared the two on real traveler questions. |
| DeepSeek | DeepSeek V4 Pro | Jun 7, 2026 | DeepSeek's free app answers with a lighter model by default; we use V4 Pro today and will realign only after a side-by-side comparison. |
 Grok | Grok 4.6 | Aug 25, 2026 | Grok 4.6 is the model Grok gives free users by default. |
 Qwen | Qwen Max | Aug 29, 2026 | Qwen Max is the model the Qwen app answers with by default. |
Data as of .
Recent changes
| Switched on | Assistant | Change | Reason |
|---|---|---|---|
| Oct 8, 2026 | Claude | Claude Haiku 4.5 → Claude Haiku 5.5 | Efficiency |
| Jul 31, 2026 | ChatGPT | GPT-5.4 mini → GPT-5.6 Luna | Upgrade |
| Jun 9, 2026 | Google AI | Gemini 2.5 Flash → Gemini 3.1 Flash-Lite | Upgrade |
| May 4, 2026 | ChatGPT | GPT-4o mini → GPT-5.4 mini | Upgrade |
What a model is
ChatGPT, Google AI, Claude and the other assistants each answer through a model: a specific version of the AI that writes the answer. Their makers release new models every few months. Sigtrip asks each assistant through one model, and your Sigtrip settings show which.
Different models give different answers, about your hotel too
Newer isn't uniformly better for a hotel. In one of our tests, a newer model wrote much shorter shortlists, so a hotel named in many answers was named in about half as many. Nothing about the hotel had changed.
Why we don't switch on release day
Providers adjust new models in their first weeks, and their apps roll them out gradually, so most travelers aren't on them yet. A same-day switch would move your score for a reason that has nothing to do with your hotel.
Since October 2026 this is a written rule for every switch we make. Before we switch, we replay real traveler questions on both models side by side, compare the difference with how much one model varies from run to run, and only then decide. Earlier switches, made before the rule was written down, didn't always follow it.
Why the model matters less than it looks
Ask the same model the same question twice and some answers already change. What earns your place is the same on every model: a clear website, strong reviews, and being mentioned on the pages assistants read.
And your position against competitors moves far less than the raw score, because they're measured on the same model.
Stability makes a trend meaningful
A trend line is only useful if the ruler stays the same. When we change a model, your trend chart shows a marker that says so, for example “Sigtrip updated the ChatGPT model”, and the score change shown on that chart is measured from the marker on, not across it.
When a provider retires a model
Providers retire models on a schedule. We record each announced retirement date and are alerted well before it, so we can choose and test a replacement in time, and the change is marked on your chart like any other.
A provider can also point a model name at a newer version without notice. Its answers then report the new version, and we detect that and mark it on your chart too.
What we don't claim
No model reached through an API is identical to the consumer app, which also uses memory, location and its own routing between models. We don't claim to show exactly what one traveler sees. We measure what each assistant's model says, the same way every day, so changes over time are real.
We show only the model that answers the traveler's question. The models we use to read and summarize answers stay internal.
Questions about the models we use? Ask the Sigtrip team.