Artificial Analysis Oct 28
A New Scale Reshuffles the Board and Widens the Gap
Intelligence Index v5 replaced a third of its public test set with private data, added FrontierCode 2.0 and Terminal-Bench-Science, and produced a new order: Claude Opus 5.5 at 58, GPT-6 Astra 57, GPT-6.1 Sol 55, Kimi K3.1 Pro 46 as the best open model, and DeepSeek V4.1 Flash with GLM-5.3 at 32 for models with published weights. The gap between the closed frontier and the best downloadable model went from 3 points to 12. Cost per completed task now spans 13x inside the same leaderboard, from GPT-6.1 Terra at $0.96 to Claude Fable 5.1 at $13.04 — which is why a single score is a worse buying signal than it was last month.
community teardown Oct 28
One Game, Two Budgets: $15 and $265
Someone rebuilt a Calvino interactive in the browser with Opus 5.5 twice and published both bills. The medium setting produced 7,746 lines for $14.98 over 53 minutes of API time and about 313K output tokens. Max with ultracode produced 31,387 lines for $264.64: 5.3M output tokens of which 3.65M were thinking, 242 web searches, 16.5 hours of API time, roughly eleven parallel agents, and a 17.7x cost ratio. The mouse input never worked in either run. As a worked example of what an agent actually costs, it is the most useful number published this week.
Nvidia Oct 28
Nvidia Opens a 4B-Active Transcription Model
Nemotron 3.5 Transcribe is a hybrid Mamba-Transformer mixture-of-experts with about 30 billion parameters total and 4 billion active per token, covering 40 languages for transcription, 31 for translation and 13 for diarization. It scores a 5.33% average word error rate on the Hugging Face Open ASR board — first among open weights — was trained on 463,062 hours of Granary v2 audio, and the quantised build transcribes 100 hours of audio in under three minutes on a single RTX PRO 6000 Blackwell.
Google Oct 29
Google Sleep: Memory That Survives Being Asleep
A two-phase framework that writes short-term memory while awake and consolidates it in a sleep phase — unlocking low-rank MoE experts, on-policy distillation, reinforcement learning on imitated trajectories, and a dreaming stage that generates synthetic data from randomly activated experts and filters it by gradient. The results claimed are unusual: near-zero forgetting on languages never seen in training, BABILong expanded to 10 million tokens, AIME 24 at 79.2% and AIME 25 at 69.0%. If it holds, persistent agent memory stops being a context-window problem.
Cartesia Oct 29
Cartesia H-Net Cuts Pretraining Tokens Tenfold
A 3.67B two-stage model with a Mamba-2 codec and a 40-layer transformer, about 7.6x compression, that reaches the final average score SmolLM3-3B reaches — using roughly 1 trillion tokens instead of 11.1 trillion, with FLOPs held within 2%. Code gains are 2.04x in-distribution and 3.10x across ten unseen languages, with a 5.37x peak on Prolog and 5.95x to 7.06x zero-shot on DNA. Learned tokenisation shifts the cost curve, not the quality curve.