Elorian Launches Glimpse, an AI That Looks Back While It Thinks

Elorian, a Palo Alto-based AI startup, has released its first model, Glimpse, as a research preview with API access. The chat interface is waitlist-only. TechCrunch

The company is led by CEO Andrew Dai, a former Google DeepMind researcher. Elorian was founded by former DeepMind researchers and came out of stealth six months ago. It raised a $55 million seed round at a $300 million valuation before it had a product. The round was co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.

Glimpse is built around access to pixels during reasoning. Most current systems encode an image once, then do the rest of their thinking — the step-by-step chain of thought — using only language tokens, the small chunks the model processes. Glimpse can return to the visual input in the middle of that process. Elorian calls the mechanism a "visual scratch space."

In practice, that means three tools are available at inference time, meaning while the model is working on an answer. Glimpse can crop into part of an image mid-reasoning. It can pull specific frames from a video mid-reasoning. It can run code to measure or compare what it sees mid-reasoning. For developers, the pattern will look familiar. It is tool use applied to vision, with cropping, frame selection and code execution offered as steps inside the reasoning loop rather than as preprocessing.

The company describes this as a change from current practice. It describes today's vision-language models as a two-step process: first translating visual inputs into language, then doing text-based reasoning. Its stated goal is to build systems that reason through the visual medium itself, the way humans do.

Training footprint is part of the pitch. Elorian says Glimpse was trained by a small team on a few hundred GPUs, the specialized chips used to train AI models. It also says the model is substantially smaller than the frontier systems it was tested against. No parameter count or training duration was disclosed in the launch materials.

Pricing is public. Glimpse costs $0.60 per million input tokens and $3 per million output tokens. Tokens are the units a model reads in and writes out, and output traces from reasoning can run long. The price is explicit.

On performance, Elorian points to its own evaluations. According to those evaluations, Glimpse outperforms GPT-6 Astra on scientific chart reasoning and on short-form video captioning. Those are vendor-reported numbers.

Initial go-to-market is narrow. Elorian is targeting robotics firms, scientific labs, and media and creative companies with lots of video. Robotics work requires precise spatial grounding, or knowing exactly where things are. Labs need measurement over figures. Video-heavy media needs frame-accurate retrieval and description.

The broader context here is inference cost versus single-pass accuracy. Revisiting high-resolution crops or video frames adds latency, or delay, and burns more tokens. The bet is that selective attention pays for itself in fewer hallucinations, or made-up details, and less manual review. In my view, that tradeoff is plausible for the three segments Elorian names, where an error in a measurement or a missed frame costs more than extra inference steps.

Looking at what this means for builders, the questions are concrete. Independent replication will matter, particularly for chart tasks where axis parsing, legend resolution and unit handling often separate demos from deployments. How stable is the crop policy under clutter and occlusion, when objects overlap or block each other. How does frame selection behave on long videos with repetitive scenes. How auditable is the code-execution step when the model measures pixels and then cites the result. Worth flagging, research preview status and a waitlisted chat interface mean most teams will meet Glimpse first through the API, where logging those intermediate visual steps will determine whether the system can be debugged in production.