It used to be so easy. If you were embedding artificial intelligence (AI) into your travel business back in 2023, the choice was almost binary: Use the newly announced ChatGPT large language model (LLM) from OpenAI or build it yourself.

The AI world has changed dramatically since then. Anthropic introduced the Claude family of LLMs, Google released Gemini and in January 2025, China’s DeepSeek upset the AI applecart by announcing its own, vastly cheaper AI model.

Stanislav Bondarenko is the CEO and co-founder of Swifty, an AI travel agent startup accelerated by Lufthansa’s Innovation Hub that books travel from a voice message prompt.  

“At the end of 2023, OpenAI was an amazing default choice; it was the most tested and reliable. The main choice was about using an existing off-the-shelf model against a model with open weights and fine-tuning,” he said.

“After that, the competition started to arrive. The landscape is now significantly wider both between providers and within every single provider.”

Use case dictates AI model

Every AI model has limits on the number of “tokens”—basic units of text such as words or punctuation—that can be input or output in a single interaction. This helps balance user access and manage costs. These limits are sometimes called the context window.

Google is known for having the longest context window. Google’s Gemini 2.5 model, announced in March 2025, ships with a 1 million token context window. If your use case involves handling vast amounts of data, then this is crucial in model choice.

A model’s “reasoning” capability is another differentiator. This is where a model breaks down a problem and shows its working, or chain of thought, before arriving at an answer. Reasoning models provide insurance against hallucinations and clarify the black-box nature of AI decisions to help engender trust in a model’s answers.

“OpenAI is still the most used and GPT-4o is the standard; they have one of the strongest reasoning models,” Bondarenko said.

AI models are increasingly being used internally, particularly for developing code. Anthropic’s Claude 3.5 and 3.7 models have won strong adoption from the developer community thanks to its “cautious” AI behavior, in that it can generate complex code through an intuitive interface but with few harmful unintended consequences.

For travel companies with a need for photorealistic image generation, such as a destination marketing organization, OpenAI’s 4o image generation and Google’s Gemini 2.0 Flash are leading the way.

Simone Lini is a former head of business development and product management at Google Travel and has recently been building Navifare, an agentic AI take on airfare metasearch.

Navifare is built on open-source Browser Use technology, which supports various LLMs.

“After reviewing benchmarks shared by Browser Use, GPT-4o emerged as the top-performing model, so that’s what I’m currently using. However, the high cost per token is prompting me to evaluate more cost-effective alternatives like Gemini 2.0 Flash,” Lini said.

Travel companies need to be open to changing models too. Lini agreed with the insights from a leaked Google memo that suggested there is no moat around models—has only been reinforced by the announcement of DeepSeek, which upset markets.

“Low switching costs combined with intense competition create an ideal environment for end-users and innovation,” he said. “The market offers numerous comparable models, and switching costs are negligible. The trend has consistently shown a steep decline in token pricing.”

Booking.com has been one of travel’s biggest proponents of AI, launching its AI Trip Planner back in 2023. This uses existing machine learning (ML) models developed in collaboration with ML scientists and leading academic institutions and is partially powered by OpenAI’s ChatGPT application programming interface and other open-source models. 

Rob Francis, Booking.com’s chief technology officer, told PhocusWire, “When we launched the AI Trip Planner, it involved combining structured data such as availability and pricing with unstructured inputs like user reviews and real-time queries. That level of integration goes far beyond just accessing a model.”

The company now works with a range of partners including OpenAI, Anthropic, AWS, Google and a variety of other open-source models, selecting the best suited for each specific use case.

Low switching costs combined with intense competition create an ideal environment for end-users and innovation. The market offers numerous comparable models, and switching costs are negligible. The trend has consistently shown a steep decline in token pricing.

Simone Lini, Navifare

Francis said, “Our choice of the right LLM for any given task is guided by principled evaluation and performance monitoring through observability. Our priority is always maintaining high standards for privacy, safety and AI ethics; we only work with models that align with these principles.”

“The implementation and integration of new models involves complexities, especially at a large scale,” Francis said. “Internally, each model undergoes a three-step review before any product testing, focusing on business use cases and legal and security reviews that ensure compliance with all relevant regulations, as well as our internal guidelines. Testing phases then entail ongoing performance monitoring and regular operational reviews. Adoption is a complex and robust process, which means integration is not entirely straightforward—but our priority in delivering value where our customers need it most, not AI for AI’s sake.”

Francis said Booking.com has built an orchestration layer to make it easier for the company’s product and engineering teams to switch easily between models.

“The orchestration layer also has enterprise capabilities to put guardrails around the conversation, allowing for moderation and to ensure we are protecting our customer’s data,” he said.

Internally, the company is using AI to improve operational efficiency. This includes Glean, a knowledge management tool with generative AI capabilities that helps unify, search and access internal data across various platforms, and Cody, which helps the company’s developers write and understand code more efficiently.

Pace of change a double-edged sword for travel companies

Navifare’s Lini said AI providers adding functionality that disrupts business models is “a universal concern.” 

Lini feels that trip planning could be “easily replicated by standard LLMs.”

“I’d focus on more operationally heavy areas or problems where the answer can’t come from LLMs,” he said. 

“My strategy is to focus on areas where differentiation is achievable, especially by leveraging proprietary insights gained from user interactions. For instance, creating a basic customer support chatbot is easy, but gaining broad adoption allows you (subject to terms and conditions) to fine-tune specialized models from user interactions, making competition from general-purpose models more challenging.”

Swifty’s Bondarenko said travel companies need to be alert when it comes to being disrupted by newer AI models.

“If your business depends on LLMs, you have to be defensible. A new model comes up every month. You always have to make sure you are on the edge of that, if that is the core of your business. The space is evolving extremely fast, and you have to have your eyes open every day for the new releases that could help or harm your business model.”