Elorian Launches Glimpse, an AI Model That Can Reason Actively with Images

Image: TechCrunch · Source

Elorian's new AI model, Glimpse, advances visual reasoning by interactively revisiting images during analysis, outperforming GPT-6 Astra on key visual tasks.

Elorian, a Palo Alto-based AI startup, has released its first AI model named Glimpse, designed to advance the capability of visual understanding beyond static description. Launched as a research preview with API access and waitlist-based chat interface, Glimpse aims to address current limitations in how AI models process images.

Unlike existing multimodal AI that processes images once and then reasons about them textually, Glimpse dynamically revisits source visuals during reasoning. It can crop specific regions, extract individual video frames, or perform measurements mid-analysis—using what Elorian calls a "visual scratch space." This mechanism is akin to how a person might zoom in on a photo repeatedly to verify their observations.

Andrew Dai, Elorian's CEO and a former researcher at Google DeepMind, points out that AI models struggle to interpret complex visual symbols and iconography found in contexts like board games or IKEA diagrams, which humans easily understand without text. This gap inspired Glimpse's approach.

Though similar interactive image processing was introduced by OpenAI in its 2025 o-series models, Elorian claims to have built Glimpse from the ground up around this capability, training it efficiently on hundreds of GPUs with a significantly smaller model size compared to leading systems.

In terms of cost, Glimpse charges $0.60 per million tokens input and $3 per million tokens output, making it more affordable than some mid-tier models but pricier than lightweight alternatives.

Elorian's evaluations indicate that Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, while remaining competitive on other visual reasoning tasks.

The company, which emerged from stealth six months ago after raising $55 million at a $300 million valuation, counts investors like Nvidia and Google chief scientist Jeff Dean. Initially, Elorian plans to focus on industries rich in video data such as robotics, scientific research, media, and creative sectors.

Looking ahead, Elorian aims to evolve Glimpse from visual comprehension to action, building models that not only see but also decide and control physical agents like robots. The CEO envisions creating "the world’s best visual reasoning model" by next year to lead across all visual reasoning benchmarks.