Elorian Glimpse: what it is and what it costs

Elorian Glimpse is the first model from Elorian, a new AI lab led by former Google DeepMind researcher Andrew Dai. It reads images and video, and it goes back to them while it reasons. It costs $0.60 per million input tokens and $3 per million output tokens, and it launched on October 16 as a research preview with API access.

The company kept its own announcement short. @ElorianAI called Glimpse "our first model, built for visual thinking" and linked to its developer platform. Dai's post, @AndrewDai, explains why he started the lab: models have handled a single still photo well for years, but they fall apart when something moves or the camera angle changes. Glimpse is his answer to that, and he claims it is competitive with OpenAI's GPT-6 Astra on the main public visual reasoning benchmarks.

How Glimpse works

Most vision models look at an image once, turn it into tokens, and then reason from that internal summary. When they miss a detail on the first pass, they can't go back for it.

Glimpse can. Elorian calls this a "visual scratch space." Partway through an answer, the model can zoom in on one part of an image, jump to a timestamp in a video, rewatch half a second of footage at a higher frame rate, or write code to measure what it sees. In one example on Elorian's launch page, the model raised its sampling from about 8 frames per second to about 14 to work out a short, ambiguous moment.

@testingcatalog summed up the early demos. In a maze test, Glimpse solved 5 of 5 mazes and Gemini 3.8 Flash solved 1. On a robot-arm planning task, Glimpse's run cost $0.035, against $0.249 for GPT-6 Astra. Elorian ran both tests itself.

TechCrunch points out that crop-and-zoom reasoning isn't new: OpenAI added it to its o-series models in 2025. The difference, per TechCrunch, is that Elorian built the whole model around it.

What Glimpse costs

The list price is $0.60 per million input tokens and $3 per million output tokens. Elorian compares that with GPT-6 Astra at $10 and $50, Claude Opus 5.5 at $4 and $20, and GPT-6.1 Sol at $2 and $10. @deedydas called it "Astra-level" for 17 times less, and that ratio holds on output tokens. He also argued that today's models are "either not capable or not affordable" for visual work at scale.

The catch is volume. Elorian admits that careful visual review "takes time, and tokens," and lists token efficiency as a top priority. A model that goes back for a second look burns more tokens than one that looks once, so the per-task number ($0.035 on the robot-arm test) is more useful than the per-token price.

Elorian's launch page says Glimpse is available through the API and a chat interface, with a "Try Glimpse free" sign-up button. I couldn't confirm whether chat is open to everyone or waitlisted, or what the free tier includes. Together AI lists Glimpse as "launching soon" for dedicated deployments.

Glimpse benchmarks: wins and losses

Elorian published its losses next to its wins, which more labs should do. In its own table, Glimpse beats GPT-6 Astra on four tests: scientific chart reading, short-form video captioning, open-vocabulary counting, and a manipulation-events test for robotics. It ties Astra on embodied spatial reasoning.

It also loses clearly on six, including dense video captioning, temporal reasoning, and long-video understanding with subtitles. On CountQA, a product-counting test, it scored 82.66 against 86.42 for both Astra and Sol. For anyone picturing an AI that counts shelf stock from a photo, that loss is worth knowing about.

Elorian ran all of these tests itself, and nobody has checked them independently yet.

My read

Glimpse looks like a credible first release from a well-funded team. Elorian raised a $55 million seed at a $300 million valuation, co-led by Striker, Menlo and Altimeter, with Nvidia and Jeff Dean also investing. For me the price is the headline. Near-frontier visual reasoning has been too expensive to run across thousands of photos or hours of footage, and $3 per million output tokens changes that math. Elorian is going after robotics labs, scientific teams and media companies first, which makes sense for a lab this size.

For a small business the uses are plainer. Think reading a gauge or an equipment nameplate from a phone photo, or checking that a job-site picture really shows the finished work. Test it on your own photos before trusting it, since every number above came from the vendor.

What this means for businesses

If image reading gets cheap and reliable, a lot of routine photo checking can be automated, from inspection pictures to proof-of-delivery shots. The model is the easy part. Someone still has to decide where the photos come from, who reviews the AI's answer, and what happens when it isn't sure. New Face Design's free process audit can help you sort out which of your photo-heavy tasks are worth handing to a model.