Elorian, founded by ex-DeepMind researcher Andrew Dai, launches Glimpse, a visual-thinking model that zooms and rewatches video while it reasons; self-reported wins over GPT-6 Astra on some benchmarks at $0.60/$3 per million tokens

Why it matters

【Key points】

Elorian, the Palo Alto lab founded by ex-Google DeepMind researcher Andrew Dai ($55M seed from Striker, Menlo, Altimeter and NVIDIA, with Jeff Dean among backers), released its first model, Glimpse, as a research preview on Oct 16. The founder's launch post went up at 17:08 UTC.

Glimpse goes back to the pixels while it reasons: it zooms into regions, rewatches short clips at higher frame rates, tracks one feature over time, runs code to measure, and can coordinate subagents. It was post-trained on an open-weights base model using a few hundred GPUs. In Elorian's own table it beats GPT-6 Astra on WGO Bench (robot manipulation events), CharXiv-R (science charts), UGC-VideoCap and FSC-147 counting, ties on ERQA, and trails on BabyVision, HSS, TOMATO and Video-MME v2. Pricing is $0.60 per million input tokens and $3 per million output, via API and a chat UI.

【Verifiable】

Blog: https://elorian.ai/blog/introducing-glimpse

Platform: https://platform.elorian.ai

Founder post: https://x.com/AndrewDai/status/2111142161584628060

【Our read】

1. A young lab with modest compute and an open base reaching near-frontier video reasoning at about 6% of Astra's price shows the bar for multimodal reasoning is dropping.

2. All scores are self-reported, the base model is undisclosed and nothing is independently replicated; "beats Astra" holds on only some benchmarks.

3. The bet is physical-world vision for robots, factories and long video. Watch for third-party evals and real deployments.