Gemini 4 Carbon is the new #1 model on the Artificial Analysis Intelligence Index with a score of 59, one point ahead of Claude Opus 5.5. It also debuts at #1 on Text Arena and #2 on Agent Arena.
Key points
Score: 59 on the Artificial Analysis Intelligence Index, top rank worldwide.
Pricing: $2 per million input tokens and $10 per million output tokens, includes a 50% launch discount.
Sources
Starlink helps ZIPAIR grow passenger revenue 12.3% YoY
Starlink helped Japan’s ZIPAIR increase passenger revenue by 12.3% year-over-year. ZIPAIR is Asia’s first airline with Starlink across its entire fleet.
Key points
Revenue: passenger revenue grew 12.3% YoY.
Deployment: ZIPAIR became Asia’s first airline with Starlink across its entire fleet.
Sources
Locus automates molecular design and outperforms SOTA
Locus automates molecular design and claims AI models that outperform state-of-the-art expert solutions. It is tested on molecular property prediction, small-molecule potency prediction with Boltz-2, and antibody design workflows using BoltzGen, Boltz-2, ESMC, and ESMFold2.
Key points
Locus develops scientific AI models that outperform SOTA expert solutions.
Locus is tested on training molecular property prediction, fine-tuning Boltz-2 for potency, and antibody design workflows with BoltzGen, Boltz-2, ESMC, and ESMFold2.
Sources
SIE Serverless launch – managed inference for small models
SIE Serverless is a managed inference product for small models. It supports 20+ task-specific workloads across Search, OCR and Agentic execution. A $500 credits launch grant is offered.
Key points
Launch: SIE Serverless – a managed inference product focused on small models
Offer: $500 credits launch grant
Scope: 20+ task-specific workloads across Search, OCR and Agentic execution
Engine: powered by an open source inference engine with thousands of GitHub stars
Sources
Locus automates molecular design, outperforms Claude Code
Locus automates molecular design and outperforms Claude Code on held-out test data. It handles molecular property prediction, potency prediction with Boltz-2, and antibody design workflows. The system proposes experiment ideas, runs several in parallel, and picks the next round based on evaluation feedback.
Key points
Covers molecular property prediction, potency prediction with Boltz-2, and antibody design workflows
Proposes experiment ideas, runs several in parallel, and selects next experiments from evaluation feedback
Sources
- Original post
- help.x.com - Linked in the post
Olmo 3 chosen by Nextcloud for AI in Assistant
Olmo 3 is selected by Nextcloud to power AI in Nextcloud Assistant and other apps. The choice supports digital sovereignty by letting users inspect model construction.
Key points
Model: Olmo 3
Application: powers AI in Nextcloud Assistant & other apps
Sources
- Original post
- allenai.org - Linked in the post
Anchor-Align improves VLA generalization with two losses
Anchor-Align introduces two simple losses on the VLM backbone to improve VLA generalization. It is presented in arXiv paper 2607.13429 and used alongside standard behavior cloning on the action head.
Key points
Anchor-Align adds two simple losses on the VLM backbone.
Standard behavior cloning is applied on the action head.
Sources
- Original post
- arxiv.org - Linked in the post
UNDERSPEC-DR framework introduced for deep research agents
UNDERSPEC-DR framework introduced for deep research agents to infer unsaid user intent. Research agents are too eager to be helpful. They’d rather spend minutes researching the wrong question than seconds asking the right one.
Key points
Research agents are too eager to be helpful
They’d rather spend minutes researching the wrong question than seconds asking the right one
Introducing UNDERSPEC-DR, a framework for deep research agents
Sources
- Original post
- arxiv.org - Linked in the post
