Elorian Glimpse visual reasoning model: $0.60 per 1M input

Elorian released Glimpse, a visual reasoning model that zooms, re-watches and measures while it thinks. The Palo Alto startup launched Glimpse on October 16 as a research preview with an API and a chat UI. The price is the hook: $0.60 per million input tokens and $3 per million output. For any business that pays people to look at photos and video, that is a number worth testing.

What actually happened

From Elorian's announcement and TechCrunch's report:

• How it works. During reasoning, Glimpse can crop and zoom into a region, re-watch a short video clip at a higher frame rate, track objects across frames, and write code to measure or compare what it sees. Elorian says the tool use was trained end to end.
• Who built it. CEO Andrew Dai is a former Google DeepMind researcher. TechCrunch reports a $55 million seed round at a $300 million valuation, co-led by Striker Venture Partners, Menlo Ventures and Altimeter, with Nvidia and Jeff Dean participating.
• Size. Elorian says a small team trained it on a few hundred GPUs, post-trained on an open-weights base model it does not name.
• Benchmarks. These are Elorian's own numbers. Glimpse scores highest or ties on six of the 12 evals it published, including open-vocabulary counting (83.36 on FSC-147) and short video captioning. It trails GPT-6 Astra on several others, and trails on CountQA product counting (82.66 vs 86.42).
• Stated limits. Careful visual review costs time and tokens. Elorian names token efficiency as a next focus. Robot control output is future work.

Why a visual reasoning model matters for your business

Most small businesses have a pile of visual work that nobody automates because it was too expensive or too unreliable. Product photos that need tagging. Shelf photos from a rep. Damage claims with pictures attached. Job-site video that a manager watches at 2x speed.

A model that can zoom into a label or count units on a shelf, at mid-tier prices, changes the math on those jobs. But "top score on six benchmarks" is not the same as "reads your SKU labels in bad warehouse light." Note that the product-counting score is one where Glimpse trails. Here is how we would test it:

• Pull 50 real images or clips from the job you want to automate, with known correct answers.
• Run Glimpse and your current model on the same set. Compare accuracy and total cost per item, not per token. Extra zooms mean extra tokens.
• Keep a human in the loop below a confidence line until the error rate is proven.
• Don't build on a research preview for anything a customer sees. Put it behind your own interface so you can swap it.

Sources: Elorian, TechCrunch.