# GPT-5.6 Built Feasible Trip Plans in 46.2% of Solvable Tasks

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/34674
Published: 2026-10-19T08:24:53.000Z
Updated: 2026-10-19T08:24:53.000Z
Section: Technology

> Travel-time tools did not prevent itinerary errors. Across 15 agents, the median success rate was 6.6% on tasks with valid solutions.

An AI travel planner can access travel-time estimates and still produce a schedule that cannot be completed. In TREK, a July 2026 benchmark of 15 agents, GPT-5.6 produced fully feasible plans for 46.2% of tasks with a valid solution, showing the difficulty of turning travel information into a workable itinerary.

TREK tested 800 trip-planning tasks. Of those, 533 had a valid solution and 267 were designed to be impossible. Agents could use tools, including one that calculated travel time between locations, and had to satisfy constraints involving budgets, trip length, business hours and travelers’ unstated preferences.

The simulated database covered 375 cities and about 210,000 records for flights, hotels, attractions and rental cars. The scoring system checked whether plans met the stated rules. Human-verified example itineraries showed that a perfect score was achievable on tasks with valid solutions.

GPT-5.6 had the highest success rate, followed by Kimi-K2.5 at 17.1% and GLM-5 at 13.1%. The median across all 15 agents was 6.6%. Claude and Gemini were not part of the test.

Travel between activities remained a common source of failure. Across the agents, 13.3% to 93.8% of tasks had too little time between same-day stops for the actual trip. GPT-5.6’s main remaining weakness was accounting for needs travelers had not explicitly stated.

Earlier tests also found limits in AI route planning, although their scores are not directly comparable because they used different tasks. In the 2024 TravelPlanner benchmark, GPT-4 produced itineraries that satisfied every requirement on 0.6% of tasks. A 2025 MapEval test covered 700 questions across 54 countries and 180 cities. None of its 30 models scored above 67%, and all scored at least 20 percentage points below human performance; distance, direction and route planning were among the weakest areas.

A 2023 analysis of Meta’s Llama-2 found that the model represented locations at several scales, including around the world, across the United States and in New York City. Researchers could recover those positions using a simple linear method. Representing where places are, however, does not itself calculate a route across roads or ensure that travel time fits between scheduled stops. Routes also depend on details such as transit schedules, transfers and business hours.

One approach is to connect models to map information. Google introduced Grounding with Google Maps in the Gemini API in October 2025, giving developers access to information about more than 250 million places, including addresses and business hours. The feature can also return distances and travel times. Google’s managed Maps Grounding Lite MCP server can search for places, check weather and calculate routes; its route options cover driving and walking.

A second approach combines language models with formal solvers. A team led by Chuchu Fan at MIT developed a system that converts a traveler’s request into mathematical constraints for a solver to arrange. It achieved a 93.9% success rate on TravelPlanner. In the same evaluation, OpenAI’s o1-preview scored 10% even when given the information needed to complete the task. The solver-based system also identified conflicts in impossible requests and helped revise them successfully in 81.6% and 91.7% of cases across two datasets.

A third approach trains models on spatial examples. MapTrace used 23,000 route samples drawn from 4,000 maps to train a multimodal model to trace paths. On MapBench, the approach raised success rates by as much as 6.4 percentage points. Related work published in 2026 tested whether models could assemble maps of simulated cities from descriptions and driving routes, and remains research-focused.

Of the three approaches, map grounding is the one already available as a commercial service. TREK’s results show that access to travel-time information alone does not guarantee a feasible schedule. Checking each leg in a map app can reveal timing problems in an AI-generated itinerary.