Elorian launches Glimpse, an AI model built to reason through images

Elorian has released Glimpse, its first AI model, in a research preview with API access and a chat interface, according to TechCrunch. The Palo Alto startup is pitching the model as a different kind of multimodal system: one that keeps inspecting visual material while it reasons, rather than treating an image as something to summarize once and then discuss in text.

That matters for teams whose work depends on charts, diagrams, video, frames, symbols, and other visual context. If Elorian’s claims hold up in real-world workflows, Glimpse could be relevant to labs, media teams, robotics companies, and creative businesses that need AI to examine visual evidence more carefully.

What Glimpse does differently

Most multimodal systems, as described by Elorian, take in an image, turn it into tokens, and then continue the reasoning process in language. Glimpse is designed to return to the source material during that process.

Elorian calls this mechanism a “visual scratch space.” In practice, the company says the model can zoom into a region, extract particular frames from video, or run code to measure and compare what it is viewing. TechCrunch compares the idea to a person checking a phone photo more closely by zooming in before making a judgment.

That distinction is important because many visual tasks are not just about naming objects. They require checking relationships, reading symbols, comparing small details, or following what changes across a clip. A chart can require numerical reasoning. A board game can require understanding icons and spatial placement. A product video can require tracking an object across time.

Elorian CEO Andrew Dai, a former Google DeepMind researcher, told TechCrunch that one sign of the problem came after he photographed a finished board game and asked a frontier model who was winning. The model could not answer, he said, because board games rely on their own visual systems of symbols and icons.

Claims against GPT-6 Astra and other models

Elorian says Glimpse performs especially well on visual reasoning evaluations. According to the company’s own testing, the model beats GPT-6 Astra on scientific chart reasoning and short-form video captioning, while producing competitive results on other visual reasoning tasks.

Those are company-reported evaluations, so buyers should treat them as a starting point rather than a final verdict. The more useful question for a small business or creative team is whether Glimpse can handle the specific visual material already sitting in their workflow: annotated diagrams, ad footage, experiment videos, storyboard frames, product photos, or training clips.

The underlying technique is not entirely new. TechCrunch notes that OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. Elorian’s argument is that Glimpse was built around that idea from the beginning, and that it did so with a smaller model and a relatively modest training setup.

The company says a small team trained Glimpse on a few hundred GPUs. It also says the model is much smaller than the frontier systems used in its benchmarks. For readers following the broader frontier-model race, the launch sits alongside other recent moves such as Meta’s Muse Wonder model family, though Elorian is focusing its first release specifically on visual reasoning.

Pricing, access, and early users

Glimpse is launching as a research preview, with both an API and a chat interface. That gives technical teams a route to test the model inside existing products, while nontechnical users can try it more directly through chat.

Elorian lists pricing at $0.60 per million input tokens and $3 per million output tokens. TechCrunch describes that as cheaper than mid-level models such as Sonnet and Sol, but more expensive than lightweight models such as Haiku and Luna.

The initial market is not casual image generation. Elorian is aiming first at organizations with large amounts of video and visual data, including:

• robotics firms
• scientific labs
• media companies
• creative companies

That positioning makes sense. A model that can repeatedly inspect frames or crop into details is likely to be more useful where visual analysis is operationally important, not just decorative. For creators and marketers, the near-term use cases may be less about making images and more about reviewing footage, interpreting visual assets, checking edits, or summarizing what happens in short clips.

What to watch next

Elorian arrived with significant investor attention before shipping a product. The startup came out of stealth six months ago after raising a $55 million seed round at a $300 million valuation, according to TechCrunch. The round was co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.

The company’s longer-term ambition goes beyond looking at images and video. Dai told TechCrunch that Elorian wants to move from visual understanding toward action: building a model of a scene and using it to decide what to do, including directly controlling robots.

For now, Glimpse should be judged on narrower questions. Can it consistently read charts better than the tools teams already use? Can it describe short videos accurately enough to reduce manual review? Can it handle niche visual systems, such as game boards, lab imagery, or production footage, without excessive prompting?

Elorian says that next year it wants to build “the world’s best visual reasoning model.” The research preview is the first test of whether that goal translates into practical advantages for teams working with visual information every day.