Elorian Glimpse: A Visual Thinking Model That Rewatches

Most multimodal AI models look at a picture once, turn it into tokens and then reason in text, a bit like describing a photo from memory. Elorian, a Palo Alto startup founded by former Google DeepMind researcher Andrew Dai, thinks that is the wrong way round. On October 16 it released Glimpse, a research-preview "visual thinking" model that keeps going back to the image or video while it reasons, zooming into regions, rewatching short intervals and running code to measure what it sees.

• Release: research preview, available October 16 through an API and a chat interface
• Pricing: $0.60 per million input tokens and $3.00 per million output tokens
• Training: post-trained on an open-weights base model by a small team using a few hundred GPUs
• Funding: Elorian raised a $55 million seed at a $300 million valuation, co-led by Striker, Menlo and Altimeter, with Nvidia and Jeff Dean participating

What Is a Visual Thinking Model?

Elorian describes Glimpse as a new category: a model that reasons about the physical world from what it sees. Mid-reasoning it can crop into part of an image, pull frames from a video at a denser rate or write code to compare measurements, a mechanism the company calls a visual scratch space. Glimpse is also trained end to end to use multimodal tools and can spawn subagents that pass text and visual artifacts to one another.

The company's own examples show why this matters. Asked whether a train is speeding up or slowing down, Glimpse tracks one window edge across the clip and notices the distance it travels per interval shrinking near the platform. In a clip of two birds leaving a door handle, it rewatches a 0.9-second stretch at about 14 frames per second to decide whether the shot cut or the camera panned. In the chat interface, users can watch that scratch space as the model works.

As TechCrunch notes, crop-and-zoom during reasoning is not brand new; OpenAI added it to its o-series models in 2025. Elorian's pitch is that it built the whole model around the idea, and did it cheaply.

How Does Glimpse Compare to GPT-6 Astra?

According to Elorian's published table, Glimpse beats GPT-6 Astra on four evaluations and ties it on one:

• CharXiv-R (scientific chart reasoning): 92.90 versus 92.10
• WGO Bench (robot manipulation events): 42.70 versus 39.42
• UGC-VideoCap (short-form video captioning): 82.66 versus 80.32
• FSC-147 (open-vocabulary counting): 83.36 versus 82.02
• ERQA (embodied spatial reasoning): tied at 83.75

Astra stays ahead on others, including BabyVision, TOMATO temporal reasoning and Video-MME v2. These are vendor-run results, so independent testing will be the real verdict. The price gap is not in dispute: Astra lists $10 input and $50 output per million tokens, and Elorian says Glimpse's rates are 94% lower. Our earlier GPT-6 Astra pricing breakdown has the full context.

Who Is Glimpse For?

Elorian is targeting organizations sitting on large amounts of visual data: robotics teams that need to identify completed actions in footage, scientific labs analyzing charts and experiments, and media companies tagging short-form video. Longer term, Dai told TechCrunch the company wants to move from seeing to doing, building scene models that can choose actions and eventually control robots. Token efficiency is the next research focus, since looking closely takes time and tokens.

For a small team with a few hundred GPUs, landing in the same conversation as frontier labs on visual reasoning is a striking debut, and a good sign for specialized model makers. More model launches live in our AI coverage.

Sources: Elorian — October 16, 2026; TechCrunch — October 16, 2026.