Elorian launches Glimpse, a visual reasoning model that zooms back into images and video

What Elorian released

Elorian, a Palo Alto startup led by chief executive and former Google DeepMind researcher Andrew Dai, released its first model, Glimpse, on 16 October 2026. The company's launch post calls it a visual thinking model: a system that reasons about the physical world from what it sees, follows objects through motion and changing viewpoints, and connects an observation to what happens next. It is labelled a research preview.

Elorian says Glimpse was built by a small team with only a few hundred GPUs and post-trained on an open-weights base model, which it does not name. According to TechCrunch, which reported the launch as an exclusive, the company raised a $55 million seed round at a $300 million valuation before shipping a product, co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.

The pitch is that today's frontier models are strong at language and code but still weak at seeing. Dai told TechCrunch that a frontier model could not say who was winning a board game from a photo, and asked why AI cannot follow wordless Ikea diagrams the way people do.

How Glimpse's visual scratch space works

Most multimodal models encode an image once into tokens and then reason in text. Glimpse is designed to keep returning to the pixels. Elorian says it can reason in visual space, revisiting and refining internal visual representations as it thinks. As an agent it is trained end to end to use tools that write code, transform images and watch videos, and it can spawn subagents that exchange text, visual artifacts and internal representations.

In the chat platform, the company shows this as a visual scratch space: the model zooms into a region, rewatches a short interval at a denser frame rate, or tracks one feature over time. One published trace follows two birds leaving a door handle. Glimpse watches the clip at 8 frames per second, then rewatches a 0.9-second stretch at about 14 frames per second to decide whether the shot is a continuous pan or a hard cut. Another follows a window edge on an arriving train from about 2.2 seconds to 5.13 seconds and concludes from shrinking displacement that the train is decelerating.

The idea is not entirely new. TechCrunch notes that OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. Elorian's claim is that it built the whole model around this behaviour, and did so cheaply.

The benchmark table, line by line

Elorian's Table 1 compares Glimpse with GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 and Gemini 3.8 Flash. It claims Glimpse exceeds Astra on four evaluations and matches it on one. The rows below are copied from the launch post; a dash means Elorian reported no score.

Pricing and how to get access

Glimpse costs $0.60 per million uncached input tokens and $3.00 per million output tokens. Elorian's table lists GPT-6 Astra at $10 and $50, GPT-6.1 Sol at $2 and $10, Claude Opus 5.5 at $4 and $20, and Gemini 3.8 Flash at introductory rates of $0.75 and $3.75. On those figures Glimpse is 94% cheaper than Astra and 20% cheaper than Flash's introductory price.

Access is through an API and a chat interface on the Elorian platform. The launch post says both are available today and invites people to try Glimpse free, while TechCrunch reported the chat interface as waitlist-only at launch, so availability may depend on when you sign up. Elorian's models page also offers volume access, custom evaluation against a customer's own data, deployment support and procurement review for companies. The company has not published rate limits or a model card.

Robotics: planning for a π0.5 policy

The most concrete demonstration is in robotics. On a LIBERO simulation task, "put the cream cheese in the bowl, then put the bowl up on top of the cabinet", Glimpse and GPT-6 Astra each acted as the high-level planner for the same vision-language-action policy, π0.5 from Physical Intelligence, which recently opened its robot models to outside developers. The planner splits the request into steps, checks the robot's cameras and decides when a step is done. Both completed the task; Elorian says Glimpse's planning cost $0.035 for the episode against $0.249 for Astra.

Elorian also shows Glimpse labelling teleoperation data, naming and boxing objects in each still and splitting an episode into six timestamped pick-and-place events, the kind of annotation robot-learning datasets need. In a recorded robot video it separates picking a marker out of a pot, ending at 8.9 seconds, from placing and releasing it, ending at 14.085 seconds. The company says direct action outputs for robot control are a next step, not something Glimpse does today.

Video archives, counting and puzzles

Beyond robotics, Elorian targets companies sitting on large video libraries. Its demos include dense captioning of a 2-minute-46-second street performance in Havana with 14 key events and a 16-second clip whose 13 events are mostly emotions and intentions. Elorian says Glimpse captured details that Gemini 3.8 Flash and GPT-6.1 Sol missed in both, though only its own checklists are shown.

Counting gets particular attention. Glimpse tracks puppies in handheld phone video so each keeps its number when the camera pans away and back, counts construction vehicles on a job site with the timestamps it relied on, and handles a still-image task that asks only for open pliers among closed pliers and lookalike scissors. Elorian also shows maze and spot-the-difference puzzles that frontier models often fumble.

These sit in a wider push on video understanding. Researchers recently showed that streaming vision-language models miss fast real-world events, and Liquid AI launched d1-ultra, a decision model that checks photos and text in one pass, a different bet on the same need for grounded visual answers.

Caveats: who ran the tests, and the HSS gap

Every number above comes from Elorian. The company states its settings, highest thinking effort and each vendor's native agent harness with a code interpreter, but notes that some rival video scores use high rather than max effort because max results were unavailable, and that HSS includes paper-reported figures. No independent lab has published Glimpse results yet.

Humanity's Sixth Sense deserves a closer look. Elorian built it with Scale AI and introduced it on 8 October as 522 questions about what a scene implies but does not show. The average person scores 93.1%. On that benchmark, Elorian's own, Glimpse scores 46.36, below GPT-6 Astra's 53.6 and Opus 5.5's 49.81. Glimpse also trails Astra on dense captioning, temporal reasoning, Video-MME v2 and product counting. That picture is closer to 'competitive and much cheaper' than 'best at vision', which is roughly how Elorian itself words it.

The worked examples on the launch post are mixed in kind. Some are recorded Glimpse traces, condensed for publication, while others, such as the tire and railway-switch examples, are explicitly reference answers rather than model output. The 'what happens next' motion clips were generated with Seedance 2.5 for illustration, and no model saw them.

Where Elorian goes next

Elorian describes Glimpse as an early step. Its stated roadmap is multimodal-native pretraining, models that both understand and generate, and direct action outputs for robots, with calibrated scene reconstruction and robot control named as work still ahead. Dai told TechCrunch the company aims to build the world's best visual reasoning model next year, one that leads on every visual reasoning test.

For now the useful questions are practical. Teams with video archives or robot data can test whether the lower price holds up on their own footage, whether the scratch-space traces make answers easier to audit, and how often Glimpse's targeted rewatching catches errors that single-pass models miss. Outside benchmark runs, a model card and published rate limits would make those comparisons much easier.