Open models cut filler from reasoning, but terser thinking is not cheaper
Newer releases from DeepSeek, GLM and Kimi use substantially less filler in their reasoning, according to a Patronus AI analysis of open models on coding tasks. The shift toward terse wording did not generally reduce reasoning-token use, however.
- DeepSeek’s filler share fell from 26.6% to 13.3%.
- Kimi K3.1 Pro had the lowest filler share at 9.8%.
- Newer releases solved 60 of 60 comparison runs.
- Only one of three model families became cheaper in reasoning tokens.
- DeepSeek’s newer model used 2.5 times its predecessor’s reasoning tokens.
Three model families became more concise
Patronus AI compared older and newer releases from three open model families across 20 agentic coding tasks. DeepSeek V4 Flash fell from 26.6% filler words in its reasoning prose to 13.3% in DeepSeek V4.1 Flash. GLM-5.2 dropped from 18.9% to 13.6% in GLM-5.3-Flash, while Kimi K3 fell from 18.0% to 9.8% in Kimi K3.1 Pro.
Across the three comparisons, filler fell by 41% on average. The newer releases solved 60 of 60 runs, compared with 58 of 60 for the older models, so the terser versions did no worse on the selected tasks, according to Patronus AI.
What Patronus counted as filler
The analysis counted five categories that a maximally terse reasoner could remove without losing information: self-talk such as “let me” and “we need,” articles, auxiliary forms of “be,” “do” and “have,” selected connectors, and first-person references such as “I” and “we.” Patronus AI excluded code blocks, shell commands and code embedded in sentences before counting English prose.
The study used 12 open models, each running once on tasks from SWE-bench Verified, SWE-bench Pro and Terminal-Bench 2.1. Runs used high reasoning effort, temperature 1.0 and the same 262k-token context cap. The selected tasks were ones the models mostly solved, with 192 of 200 current-model runs succeeding.
Among the current open models examined, Kimi K3.1 Pro had the lowest filler share at 9.8%. Qwen3.8 Flash had the highest at 23.7%.
Shorter reasoning did not usually lower token use
Patronus AI found that only one of the three model families became cheaper in reasoning tokens. DeepSeek’s newer model used 2.5 times the reasoning tokens of its predecessor, despite its lower filler share. The finding separates concise wording from lower computational or billing costs: removing articles and other redundant language does not necessarily make a reasoning trace shorter in tokens.
The results matter because reasoning text is billed, logged and increasingly read by monitors, according to Patronus AI. For users and developers evaluating open models, the study suggests that a cleaner-looking trace may improve readability without guaranteeing lower usage costs. Patronus AI has published each run, including its full trajectory and score, in the PatronusAI/few-word-do-trick-traces dataset on Hugging Face.
Patronus AI: Few Word Do Trick: Open Models Cut the Filler | Patronus AI
