Relay-Bench: Even GPT-5.5 Scores Just 43.3% on a Multi-Domain Reasoning Test

In July 2026, researchers published Relay-Bench on arXiv, a benchmark that tests whether language models can solve linked tasks from different subject areas in a single prompt. The best-performing model tested, GPT-5.5 in xHigh mode, scored just 43.3%. Even the strongest system failed to solve half of the tasks.

What Relay-Bench Measures

Relay-Bench evaluates whether a model can keep several reasoning domains in play within one prompt. The entire test set consists of composite tasks: several single-domain subtasks are chained together into one sequence that cannot be solved by reasoning in a single field alone. Each task contains from 2 to 13 subtasks and stays fully text-based. No multimodal input or output is required, which makes the test easy to reproduce on any model.

The authors call the benchmark "unsaturated": current models still have enormous room to improve, the ceiling is far away, and no single update will close the gap. That sets Relay-Bench apart from many older tests, where top models long ago hit 90–100% and stopped being distinguishable from one another.

• The leaderboard is topped by GPT-5.5 (xHigh) with 43.3%
• Each task is stitched together from 2–13 linked subtasks
• Domains: visual reasoning, code, math, information retrieval (web search), problem solving, general knowledge, data analysis
• Text-only format, with no images in input or output
• Models are allowed to run code, search the web and use any tools

Why Even GPT-5.5 Scored Only 43%

The tasks are deliberately made harder, which is why scores are low even for the strongest models. According to the arXiv description, layers of difficulty are added to the base chains through prompt encoding and deliberate context bloat. The model has to filter out the noise without losing the thread between domains. There are no restrictions outside the model harness: the authors explicitly encourage code execution, web search and all available tools, yet even so the 43.3% bar remains unbroken.

The key problem is switching between domains. A model may confidently solve a math subtask and a code subtask separately, but it stumbles when the result of one subtask has to be passed into the next within a single chain of reasoning. This "relay" is what gave the benchmark its name.

"An unsaturated, holistic, text-only benchmark measuring the ability of LLMs to complete a set of tasks from different domains in a single prompt," from the

Relay-Bench description on arXiv.

How This Differs From Conventional Tests

Relay-Bench tests not isolated skills but their combination within one task. Classic benchmarks measure code, math or general knowledge separately. Here, seven domains are woven into a single chain, and a failure at any link sinks the whole task. This format is closer to real working scenarios, where an agent has to search for data on the web, do calculations and write code within a single assignment.

What This Means

Relay-Bench shows the gap between models' narrow skills and end-to-end reasoning: LLMs can handle each domain piece by piece, but they cannot yet link them into a long multi-domain chain. The room for growth is huge, since 43.3% is still far from the ceiling, and it is precisely such "unsaturated" tests that will set the bar for the next generations of models and agent systems.

Frequently Asked Questions

What is Relay-Bench?

Relay-Bench is a text-only benchmark presented on arXiv in July 2026. It tests how language models solve composite tasks spanning seven domains (code, math, web search, data analysis and others), assembled into a single chain within one prompt.

Which model scored the best?

The best result belongs to GPT-5.5 in xHigh mode: 43.3%. That means even the strongest model tested solved fewer than half of the tasks, and the benchmark remains "unsaturated" and far from its ceiling.