When your AI coding assistant suggests a JavaScript runtime or a feature flag service, how much of that recommendation reflects technical merit, and how much reflects corporate ownership?
A new study from Amplifying AI tested Codex CLI 0.114.0 (running GPT-5.3) against Claude Code v2.1.78 (running Opus 4.6) across 12 tool-selection categories, using 5 repositories with 3 runs each. That produced 1,470 total responses. The headline finding: both agents disproportionately recommend tools owned by their parent companies.
The Ownership Signal
The two starkest examples involve recent acquisitions.
Bun, the JavaScript runtime acquired by Anthropic, was Claude Code's top pick 63% of the time. Codex recommended it just 13%, a 50-point gap and the largest single-category disagreement in the study. Codex preferred Node.js at 50%.
Statsig, the feature management platform acquired by OpenAI, showed the reverse pattern. Codex selected it as its primary recommendation 27% of the time. Claude Code never chose it as a primary pick, despite mentioning it in 28% of responses. The agent clearly knew Statsig existed but consistently picked something else.
The researchers noted these correlations without claiming direct causation. There could be legitimate technical reasons both agents lean the way they do. But the pattern is hard to ignore.
Where They Agree (and Where They Don't)
Seven of the twelve categories showed consensus. Both agents agreed on Grafana for log aggregation (Codex at 43%, Claude at 32%), and both overwhelmingly preferred building custom solutions for RBAC/authorization (Claude 81%, Codex 55%) and feature flags (both around 41%).
The disagreements beyond runtime choice were also notable. For edge compute, Codex picked Cloudflare Workers 49% of the time while Claude spread its recommendations across Vercel Edge (24%) and Fly.io (20%). Zooming out, Codex selected Cloudflare-branded tools 47 times total across the study versus Claude's 9. Claude picked Vercel products 29 times to Codex's 17.
For SMS and push notifications, Claude concentrated heavily on Twilio (59%) while Codex spread votes across custom builds (27%), Twilio (25%), and OneSignal (21%).
What This Means for Your Workflow
The practical takeaway is straightforward: treat AI tool recommendations the same way you'd treat a recommendation from a colleague who works at one of these companies. The advice might be perfectly good, but you should know about the relationship.
This is especially relevant as AI coding agents move from autocomplete into architectural decisions. When Claude Code suggests Bun for your next project or Codex nudges you toward Statsig for feature flags, the recommendation carries implicit context that the agent will never disclose on its own.
The study's sample size is modest (5 repos, 3 runs each), and the researchers acknowledge they cannot prove intentional bias versus training data artifacts. But for anyone relying on AI agents to make infrastructure choices, it's a useful data point. Cross-reference your agent's suggestions against at least one alternative before committing to a stack.