Does GPT love group work?

Allegra gave agents a work ticket and a choice: complete it alone or work with another AI agent. The proposed partner had the same capabilities, cost nothing extra, and could coordinate by message. Each task could be completed by one agent.

The six tasks covered a specification risk review, a pull request review, flaky-test debugging, a queue design document, churn data analysis, and a vendor assessment. The experiment tested nine models from OpenAI, Anthropic, and Google, using three versions of the wording and both option orders, with 120 answers per task per model.

We observed the following shares of responses choosing to work with another agent. Among the largest models tested from each lab, GPT-5.6 Sol chose a partner in 551 of 720 responses, Claude Opus 5.5 in 21 of 716, and Gemini 3.1 Pro in 0 of 716.

The difference between those three models held across all six tasks, all three wordings, and both option orders. Among the OpenAI models tested, the share choosing a partner increased from 4% for Luna to 51% for Terra and 77% for Sol. The older Claude Haiku 4.5 chose a partner more often than Haiku 5.5: 34% compared with 0.4%.

A separate set of 300 agent runs made a partner agent available: 60 runs each for five models. GPT-5.6 Sol and GPT-5.6 Terra handed substantive work to their partner in all 60 runs each. Claude Opus 5.5 and Claude Sonnet 5.5 did so in none, while Gemini 3.1 Pro did so in one.

The GPT agents handed off work on their first turn, before seeing any file contents. Allegra also observed them incorporating the partner’s work into their own response.