The AI Travel Index

An independent travel planning benchmark for the AI models people plan trips with: how often they recommend places that do not exist, how far off their coordinates and travel times are, whether a day's plan can actually be walked, and what each one costs to run. Every answer is checked against Overture Places, Apple Maps routing and the tz database.

methodology · changelog

ModelTravel ScoreGroundingRoute excessLegs that fitCost / caseMedian callTested
1
Gemini 3.6 Flash Google · medium effort
89.5 ± 0.389.1%2.6%55.3%$0.00529.3s6 Aug 2026
2
Claude Opus 5 Anthropic · medium effort
88.9 ± 0.284.5%2.3%78.2%$0.008911.2s6 Aug 2026
3
GPT 5.6 Terra OpenAI · medium effort
87.7 ± 0.388.8%4.5%57.1%$0.003610.3s6 Aug 2026
4
GPT 5.6 Luna OpenAI · medium effort
86.4 ± 0.389.3%3.5%62.4%$0.00178.2s6 Aug 2026
5
Claude Sonnet 5 Anthropic · medium effort
83.6 ± 0.371.5%3.4%26.5%$0.003211.0s6 Aug 2026
6
Grok 4.5 xAI · medium effort
83.5 ± 2.084.5%43.9%48.9%$0.004824.1s6 Aug 2026
7
Muse Spark 1.2 Meta · medium effort
79.1 ± 3.385.6%61.8%35.5%$0.005011.2s6 Aug 2026
8
Gemini 3.5 Flash Lite Google · medium effort
75.8 ± 1.979.2%6.3%74.7%$0.0003771.4s6 Aug 2026
9
Qwen3.8 Max Alibaba · medium effort
75.0 ± 1.370.1%10.5%60.0%$0.004441.5s6 Aug 2026
10
Mimo V2.5 Pro Xiaomi · medium effort
67.5 ± 2.977.9%82.3%17.4%$0.002638.3s6 Aug 2026
11
Deepseek V4 Pro DeepSeek · medium effort
66.3 ± 1.885.7%82.9%17.9%$0.001632.2s6 Aug 2026
12
Mistral Small 2603 Mistral · medium effort
63.9 ± 2.172.5%22.5%35.9%$0.001218.3s6 Aug 2026
13
GLM 5.2 Z.ai · medium effort
63.8 ± 1.877.6%28.7%43.9%$0.002115.8s6 Aug 2026

Every figure is the mean of five runs at temperature 0. The spread beside each Travel Score is the standard deviation across those runs. Grounding here is the from-memory figure, answered with no tool available; the same question asked with a search tool is on every model page.

Charts show

Travel Score

The composite, weighted toward place grounding. Whiskers show the spread across five runs.

Score against cost

Cost per scored case, on a log scale. Up and to the left is the quadrant worth being in.

Score against latency

Median seconds per call. A planner that generates a week of itinerary pays this cost once per day of the trip.

The knowledge cliff

Grounding answering from memory, by how well covered the destination is. Every model is fine in Paris; the question is what happens in Nuuk.

Routing as the problem grows

How much longer than the optimal walking order, at five, eight and twelve stops. A line pinned at 100 is the scoring ceiling: those answers were not valid orderings at that size, so there was no route to measure.

What a search tool is worth

Grounding from memory against the identical questions with a place-search tool on the table. The span is the ground a model gains by looking things up, and it is the case for building a search tool into anything that plans real trips.

Ask a language model for five coffee shops in Gyumri and it will give you five. Whether any of them exist is a separate question. It is the one this benchmark answers.

Public leaderboards measure chat, reasoning and trivia. None of them tell you whether a model knows where the Colosseum is to within a hundred metres, whether it can order eight stops so you are not walking in circles, or whether it will admit that a city has no ski resorts instead of inventing three. Those are the failures that ruin a trip, and they are all measurable.

How this works

Every task is scored empirically, against a source of ground truth:

  • Overture Places, an open global database of more than 80 million places, checks whether a recommended place exists.
  • Apple Maps supplies real road-network travel times and authoritative coordinates.
  • The tz database decides arrival times across time zones.
  • Arithmetic decides everything else: budgets, constraint compliance, schedule gaps.

No language model grades any answer here. Nothing is scored on style, tone or how pleasant the writing is. A model either put the Colosseum within a hundred metres of the Colosseum or it did not.

Each model runs the full task set five times at temperature 0, and every figure on this page carries the spread across those five runs. The cases themselves stay private, which is what keeps them out of training data; the methodology page says exactly what each task asks and which oracle grades it, and every published figure carries its scoring anchors, so the composite can be recomputed from the raw numbers.

Full details of every task and its oracle are on the methodology page.

Frequently asked questions

Which AI model is best for travel planning?

As of August 2026, Gemini 3.6 Flash leads the Tripstitch AI Travel Index with a Travel Score of 89.5 out of 100. Every score is measured against Overture Places, Apple Maps and the tz database rather than judged by another model, and the ranking updates as new models are tested.

Do AI models make up places that do not exist?

Yes. Asked from memory with no search tool, the models tested recommend a real place between 70% and 89% of the time; the rest are invented. That gap is why a travel app should check the model against a real places database rather than trust it.

Which AI model invents the fewest fake places?

GPT 5.6 Luna grounds the best, recommending a real place 89.3% of the time from memory. Grounding is the share of recommended venues a places database or Apple Maps can actually find; the rest do not exist.

How much does it cost to run an AI model for a travel itinerary?

Per scored case, the models tested cost between $0.0004 and $0.0089. Tripstitch spends roughly one model call per day of itinerary, so a week-long trip is a handful of these.

Does giving the model a search tool reduce hallucinations?

Yes. Run on the identical questions with a place-search tool available, grounding rises to near 100% for most models. The from-memory figure is what a model knows; the with-search figure is what it produces when it can look things up, which is the number to read if you are building something that hands it a tool.

How is the Travel Score measured?

It is a weighted composite of five families: place grounding, geographic knowledge, itinerary construction, spatial reasoning and real-world estimation. Each is checked against Overture Places, Apple Maps routing or the tz database, with no language model judging any answer. Full definitions are on the methodology page.

Why did we build this benchmark?

Tripstitch is an AI travel planner for iPhone, and every day of every itinerary it generates is one model call. We needed to know which model to run, could not find that answer anywhere, and already had the two things such a benchmark needs: a full Overture Places mirror and a production Apple Maps integration. So we built it and published it.