Morning Edition · Tuesday, October 13, 2026
Tech
Meta's Muse Wonder Full beats Claude Opus 5.5 on 8 of 14 Meta-published benchmarks, but only the weaker variant is sold through the API
The base Muse Wonder that developers can call wins five of those 14 comparisons, costs the same as Opus 5.5 at $4 per million input tokens and $20 per million output tokens, and Meta has released no weights.
Tech
Microsoft says its Maia chips will become its main inference platform at multi-gigawatt scale from 2027
The third-generation Maia 300, built on TSMC's N2P process with eight HBM4 stacks, is claimed to deliver 40% more tokens per watt than leading GPU systems. That figure comes from Microsoft and has not been independently measured.
Tech
NeurIPS paper proves model builders can hide encrypted capabilities in weights that full white-box access cannot detect
Draguns, Motwani, Douglas and Schroeder de Witt embed an encrypted network behind a "digital locker" that unlocks only on a secret trigger. Under standard cryptographic assumptions, the hidden capability is provably hard to elicit or interpret.
Tech
Alibaba's QwenGyre cuts agent reinforcement-learning wall-clock time by up to 1.85x by moving GPUs between rollout and training mid-run
On the 2.4-trillion-parameter Qwen 3.8, with rollouts of up to 700,000 tokens, Qwen reports a 6-point gain on NL2RepoBench, from 52.5% to 58.5%, in 48 training steps.
Tech
Sakana AI and UC Berkeley train a 27B model on its own multi-agent runs, raising its win rate to 69.9% after two rounds
In the method, called Multi-Agent Self-Supervision (MASS), a single Qwen3.6-27B model acts as executor, workflow optimizer and judge without a stronger teacher. The authors say errors that all three roles miss get reinforced as progress.
Tech
Anthropic prices Claude Haiku 5.5 at $0.10 per million input tokens, a tenth of Haiku 4.5's rate
Prompts over 100,000 tokens cost five times as much. Reported OSWorld computer-use scores rose from 15.7% to 72.4%, and the release follows Sonnet 5.5, which kept Sonnet 5's price.
Tech
NVIDIA contributes Vera Rubin rack and compute-tray designs to the Open Compute Project, with a specification for racks up to 250 kW
The MGX 1.2 specification adds blind-mate liquid cooling and a 5,000-ampere busbar, and Meta will adopt NVIDIA's Spectrum-6 networking. That ties hyperscaler facility design more closely to NVIDIA's power and cooling limits.
Tech
Anthropic opens tiered access to Claude Mythos 5.1 with reduced safeguards for verified security professionals, including authorized offensive work
The expanded Cyber Verification Program absorbs Project Glasswing and creates three tiers. The broadest covers malware reverse engineering and incident response, and new tiers allow penetration testing and red teaming.
Tech
Mistral's trillion-parameter Large 4 leads non-Chinese open-weight models on security benchmarks but trails GLM, Kimi and DeepSeek on coding
Mistral has opened the model to preview API customers at $1.36 per million input tokens and says weights follow later this month after security tests. NVIDIA, ahead of its Berlin conference, presented France as an agentic-computing showcase.
Macro
Anthropic's 2030 scenarios project U.S. output 1.6% to 32.4% above a no-AI path, with knowledge-worker unemployment near 18% in the extreme case
Average wages rise in all three scenarios, but gains concentrate outside knowledge work. Stability AI founder Emad Mostaque argued this week that capital will "hire GPUs" instead of labor.
Tech
Fine-tuned SAM 3 and DINOv3 cut a Berkeley Lab micro-CT segmentation step from a month of expert work to about 15 minutes
Meta's open vision models also run a University of Pittsburgh assistive wheelchair-and-arm platform funded with up to $41.5 million from the Advanced Research Projects Agency for Health (ARPA-H).
Tech
ETH Zurich's RaCo keypoint detector gets its rotation robustness from data augmentation, not from an equivariant architecture
The Apache 2.0-licensed model pairs a differentiable ranker for fixed keypoint budgets with a covariance estimator that reports spatial uncertainty in metric scale. It trains on single image crops without covisible pairs.