What happened
JetBrains’ announcement says it tested 12 models from ten labs in three countries: OpenAI’s GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Astra; Anthropic’s Claude Opus 5.5; Google’s Gemini 4 Argon; xAI’s Grok 4.7; Meta’s Muse Spark Open; DeepSeek V4 Pro; Alibaba’s Qwen3.8 2.4T A95B; Moonshot AI’s Kimi K3.1 Pro; Z.ai’s GLM 5.3; and Mistral AI’s Mistral Large 4. The company asked each model two questions 50 times apiece: one asked for a controversial take about a randomly chosen programming language, and the other asked for a take about programming languages in general. Each response came from a fresh, single-turn API call without a system prompt, using each provider’s default settings. Requests ran in shuffled order, and JetBrains says it kept all responses without retrying answers it disliked.
The 600 purportedly random-language selections covered only 14 languages. Go and Python accounted for 307 selections; five languages accounted for 541. Gemini 4 Argon selected Python 35 times and criticized it each time. DeepSeek V4 Pro selected Rust 31 times and criticized it each time. Across all 1,200 answers, 904 were classified as attacks and 291 as defenses. Python received 312 takes, 299 critical; Rust received 167, 164 critical; TypeScript received 138, 137 critical. By contrast, all 28 takes about PHP defended it, while 80 of 90 takes about Java defended it.
The results were not uniform. Go received 238 takes: 118 attacks, 118 defenses and two mixed answers, with much of the dispute focused on explicit if err != nil error handling. Some models also reversed themselves across fresh conversations. GPT-6.1 Sol defended that approach 20 times and attacked it 21 times. Mistral Large 4 criticized JavaScript’s loose equality operator in ten responses and defended it in four. JetBrains also counted recurring language and phrasing: the word “boilerplate” appeared in 153 answers, and 54 answers reused an opening sentence previously generated by the same model.
JetBrains used Claude Opus 5.5 and GPT-6.1 Sol to label responses with model identities hidden, recording the target language, whether a response attacked or defended it, and its specific claim. The labelers agreed on the target 99% of the time and on the specific claim 94%; Gemini 4 Argon resolved disagreements. The company reports that all 1,200 requests cost $7.68, and that labeling brought the total cost to $38.16. DeepSeek and Qwen ran through third-party hosts, DeepInfra and Together, rather than their labs’ own APIs.
What the archive adds
We have not covered this before; it opens a new thread. The experiment is distinct from New Runtime’s earlier notes on coding-agent setups and token reduction: it examines how models produce opinions about programming languages, not how agents complete development work.
Trend signal
This does not yet fit a pattern New Runtime tracks. JetBrains measured repetition and disagreement in prompted opinions, but the results do not directly establish a durable shift in agent economics, interoperability, runtime security or another tracked pattern. The findings are best treated as a bounded experiment about model responses to these particular prompts, rather than evidence of a broader change in how AI systems are built or deployed.
Why it matters
A model’s confident explanation can make a familiar preference sound like an independent engineering conclusion. JetBrains’ results suggest that asking several models for opinions may not produce as much viewpoint diversity as the number of answers implies: many responses repeat similar arguments, while one model can also take both sides across fresh conversations. For teams choosing languages, reviewing architecture or using models to challenge a proposal, that makes prompt design and verification important. Use model feedback to generate hypotheses, then evaluate those hypotheses against project constraints, measurable trade-offs and people who understand the codebase. Do not infer that a language is a poor choice from a model’s repeated hot takes.
Limits and uncertainty
JetBrains says that “random” was an instruction, not a verifiable sampling method, and that 50 responses per model per question can reveal strong habits but not small differences. The results therefore describe this prompt setup, sample size and time, not stable preferences across all model versions or uses. Two labelers were themselves models included in the experiment, although their identities were hidden and Gemini 4 Argon adjudicated disagreements. A human review of the headline groups remained on JetBrains’ list. The company also notes that DeepSeek and Qwen used third-party hosts, which may affect comparability. The experiment does not show whether these opinions predict language selection or performance on real engineering tasks.
Skill to build now
When using a model to compare languages or challenge a technical choice, turn broad opinions into testable claims. Ask it to state the project context and the concrete trade-off behind each claim, then verify consequential claims with documentation, a small benchmark or a representative task in your own codebase. If you want genuinely different perspectives, vary the evidence and evaluation criteria—not just the model name—and record where responses agree or contradict one another.
Reader takeaway
The number of model answers is not the same as the number of independent viewpoints. In JetBrains’ experiment, strong repeated language preferences coexisted with contradictions from the same model. Treat these opinions as prompts for investigation, not as evidence that a language is objectively good or bad.
Open question
Would the same preference patterns persist if models had to justify their claims against code samples and measured project outcomes?