Elorian Launches Glimpse, a Visual Model for "Workspace"

Key Takeaways

1Elorian launches Glimpse, a visual reasoning model that refers back to images during inference

2The model is accessible via API, with a chat interface on a waiting list and public pricing

3Glimpse claims to outperform GPT-6 Astra on certain visual tests, according to Elorian

4The company initially targets video applications and ultimately aims for robot control

Why it matters : Glimpse offers a different approach to automated visual reasoning, with clear ambitions in the applied AI market for video and robotics.

Elorian introduces Glimpse, a first model focused on visual reasoning that refers back to images during inference. The company displays its pricing, claims results against benchmarks, and aims for immediate video applications, before an announced transition from vision to action, including robotics.

Claimed performance against GPT-6 Astra

According to Elorian's evaluations, Glimpse outperforms GPT-6 Astra in reasoning about scientific graphs and in captioning short videos. The company also reports competitive results on other visual reasoning tasks. It notes that the model is substantially smaller than the leading systems it is compared to.

API access, chat waitlist, and public pricing

Glimpse is available in research preview via an API, while the chat interface is currently on a waitlist. The model is priced at $0.60 per million input tokens and $3 per million output tokens. These prices are stated to be lower than those of mid-tier models like Sonnet and Sol, and higher than those of lightweight models such as Haiku and Luna.

A model built around a "visual workspace"

During its reasoning process, Glimpse returns to the original image, can select a portion, isolate specific images from a video, and execute code to perform measurements or comparisons on what it analyzes. According to Elorian, this system constitutes a "visual workspace." This approach differs from the common practice of converting images into tokens and then reasoning through text. OpenAI had added cropping and zooming during reasoning to models in the o series in 2025, while Elorian claims to have structured the entire model around this principle, with training conducted by a small team on a few hundred GPUs.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Context, launch, and target sectors

Elorian released its first model, Glimpse, on Friday, presented today as a scorecard and a thesis on the vision of AI. Based in Palo Alto, the company raised $55 million at a valuation of $300 million before any product, with a round announced six months ago co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Jeff Dean. The initial target applications involve organizations handling numerous videos, including robotics, scientific laboratories, media, and creative industries. In the long term, Elorian plans to transition from vision to action, including direct control of robots, and Andrew Dai aims for "the best visual reasoning model in the world" next year. CEO of Elorian and former researcher at Google DeepMind, Andrew Dai illustrates perceived limitations through a personal experience with board games and discusses the difficulty models have in interpreting iconography, while humans assemble furniture from diagrams. Elorian argues that current models do not truly see.