syuan03/benchboard is a repository that collects benchmark scores for publicly available AI models together with source links. In the October 24, 2026 commit be56c57, it newly included voice benchmark results for Grok Voice 3.0, Gemini 3.8 Live, and Gemini 3.8 Live Extended Thinking. The changes amount to 208 changed lines in total (199 insertions, 9 deletions), covering the README statistics wording and the statistics SVG, two data pack files, two audit documents, index.html, and the data loader.
The premise of this repository is the context behind the scores rather than the scores themselves. The README states that it attaches the model release, benchmark variant, evaluation setup, and original source to every score, and specifies that results from different harnesses or versions are kept separate. The explanation that it does not create cross-benchmark composite scores also remains in place. This time, the Grok entry separately includes 10 results drawn from xAI's Grok Voice 3.0 release and the Artificial Analysis speech-to-speech leaderboard, and manages audio quality, reasoning, agentic tasks, conversational dynamics, preference, EVA-Bench accuracy, latency, and input audio cost as distinct metrics. The October 13 release Elo and the October 24 live Elo are likewise kept as separate entries, and on the Gemini side, 8 results from the Google announcement and the DeepMind evaluation chart are divided into audio quality, agentic voice performance, Sierra banking, Big Bench Audio, and measured input audio cost, while model IDs and modality limitations follow the model card.
However, the 'verified' in the commit message means that the repository manages scores by attaching sources — it does not mean that it reproduced the scores itself. The dates on the evidence documents left in the data pack, September 15, 2026 and September 2026, precede the commit date. The repository statistics increased from 200 to 203 models and versions, from 582 to 585 benchmark families, from 9,565 to 9,584 public results, from 313 to 319 primary sources, and from 137 to 140 edited releases. This can be read as a sign that voice model comparison has joined the data axis, but since evaluation conditions differ by metric, it would be premature to judge which side is superior on the basis of this repository alone.