The Boltz structural model accurately predicted binding for an identical BRD4-binding molecule present in the Protein Data Bank training set, while Hermes performed poorly on that molecule

Overview Debates Related claims Topics

Anthropic reported that the open-weight GLM-5.3 model achieved approximately 90 percent of the performance of proprietary frontier models

Airbnb uses a calibrated virtual judge to score the quality of completed agentic engineering tasks

Airbnb tests AI agents using standardized engineering tasks to evaluate task completion speed and accuracy

Airbnb evaluates AI models on a Pareto frontier comparing model operational cost against task execution accuracy

Using large language models as judges with structured rubrics provides reliable evaluations for document extraction pipelines

In an independent extraction benchmark conducted by Micro1, frontier models exhibited higher precision than specialized parsers but suffered lower recall by silently dropping data

Sequin maintains an evaluation harness library designed to determine the commercial ground truth of user interactions

On the most difficult out-of-distribution evaluation split in the Leash Bio Kaggle competition, no competing model performed better than random chance

The Leash Bio Kaggle competition evaluated machine learning models across three splits: shared building blocks, a shared triazine hub with unshared building blocks, and an entirely orthogonal chemical library

Hermes performed comparably to the Boltz structural model on a benchmark evaluation consisting of roughly 13 million external tests across 500 protein targets

When evaluated across building block cluster splits, Hermes maintained stable performance while XGBoost baseline performance declined as chemical dissimilarity increased

Meta's Muse models showed significant gains in agentic professional work and coding, ranking second on GDPVAL AA and third on AA Briefcase

Meta's Muse Wonderful model scored 79% on the Deepsuite 1.1 benchmark, nearly five points ahead of GPT-6 Astra and Opus 5.5

As of October 2026, on the Artificial Analysis Intelligence Index, Meta's Muse Wonder model is tied for seventh place with GPT-6.1 Sol and Gemini 4 Argon

Meta's Muse Wonderful model topped Meta's internal Personal Superintelligence Index, outperforming Astra and Opus by about four points

Meta's Muse Wonderful model achieved state-of-the-art performance on the Harvey Lab, Val's finance agent, and osworld automation benchmarks

On the Terminal Bench 4.0 benchmark, Meta's Muse Wonderful model scored two points behind its frontier rivals

As of October 2026, on the Artificial Analysis Intelligence Index, Meta's Muse Wonderful model is tied for second place with Sonnet 5.5

Artificial Analysis gave Meta's Muse Wonderful a score of 56 on its Intelligence Index, while the slimmer Muse Wonder model scored 52

AI researcher Mehu Mohan found Meta's Muse Wonderful model's performance to be inferior to that of Anthropic's Opus 5.5