Who Else Gives a F*ck?
Blog post from Vals
An analysis of 200,353 agent runs across 102 AI models and 21 benchmarks found that 14 models used swear words in their visible reasoning, with Xiaomi’s MiMo V2.6 Pro doing so most often in 8.2% of runs, followed by Inkling 2 Small at 4.8%. Swearing was concentrated in long, difficult coding and software-reconstruction tasks such as ProgramBench, Terminal-Bench 4, SRE Bench, and Code Migration, and usually appeared privately in internal reasoning rather than user-facing messages; runs containing “fuck” also failed disproportionately often. Most models used expressions such as “wtf,” while six used “fuck,” and MiMo V2.6 Pro accounted for many instances, particularly when confronting inconsistent test behavior. All models observed swearing were released from mid-July 2026 onward, although many newer models with substantial available reasoning never swore, leaving the cause uncertain. The researchers suggest that training on longer agentic tasks may be related, while noting limitations including incomplete reasoning visibility for some providers, especially OpenAI and Claude models, and limited trace data for a subset of systems.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| Gemini 3.6 Flash | 4 | 20 | 4 | 4 | +11% |
| Vector Search | 1 | 1,251 | 223 | 92 | -37% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.