Cheap models beat frontier models at writing like a human

The proposal. Louis-François Bouchard, founder of Towards AI, released Tone-Bench, a benchmark built to quantify the "slop" problem in AI-generated writing. Where existing checks like EQ-Bench's slop score just count banned words, Tone-Bench scores whether a model can write a script that sounds like a specific person, namely Bouchard himself, without cringing. This is not a generic writing-quality test; it is a voice-matching benchmark with ten scripts, detailed rubrics, and thousands of dollars in evaluation runs behind it.

Why now. The immediate catalyst is Andrej Karpathy's "slop apocalypse" prediction for 2026, which Bouchard says has in effect already arrived on social media. The deeper driver is practical: Towards AI produces thousands of educational articles and hundreds of videos with a small team, and the organization spends real effort cleaning up AI-generated drafts. The benchmark is as much an internal content-production tool as a public leaderboard.

The method. Each model sees a roughly 14,000-token prompt containing a content brief, research materials, examples of Bouchard's writing, and stylistic instructions. The model generates five drafts of ten scripts. Four LLM judges score each draft on nine metrics grouped around tone, writing quality, instruction following, research accuracy, and slop. The multi-draft design deliberately catches intra-model variability, since a model like Grok 4.7 can produce one Opus-quality draft out of five, which becomes a misleading signal from a single sample.

The money. Building the benchmark was far from cheap: roughly 500 hours of human work, four months of high-tier API subscriptions, and repeated runs that cost thousands of dollars. A single change to the judge rubric costs over $2,000 to re-run; changing one line in the generation prompt costs over $4,000. This is a useful warning for teams building their own preference or voice-evaluation pipelines.

The findings. The headline result is practical, not poetic: for most teams, the best writing model is not the best model, but the least annoying one. GLM 5.3 Flash scores nearly as well as Opus 5.5 Extra High at roughly 1/50th the cost. Bouchard also observes a clear family bias in LLM judges, with OpenAI models consistently preferring their own outputs, and reports that some judge models seem trained to dislike Anthropic or OpenAI rivals.

The trade-off. The benchmark is explicitly subjective, measuring "can this model write like Louis-François," not "can this model write well." The practical advice is more transferable: make prompts long, include positive and negative examples, use one high-thinking judge pass instead of repeated loops, generate many cheap drafts and merge the best parts with a smart judge, and always keep a human reviewer. The benchmark itself is a helpful signal, not a universal verdict.