Elorian Glimpse: $55M Startup Ships a Model That Rewatches the Clip

Elorian Glimpse is the first model from the Palo Alto startup that raised $55 million before shipping anything, and it makes a specific bet: AI should keep looking at an image or video while it reasons, not glance once and think in text. It launched October 16 as a research preview with an API and chat app, priced at $0.60 per million input tokens and $3 per million output.

What actually happened

Elorian came out of stealth about six months ago and released Glimpse on Friday. TechCrunch had the exclusive, and the company published a long launch post with recorded traces and benchmark tables the same day. Glimpse is available through an API and a chat interface; the company calls it a preview.

The core pitch is what Elorian calls a "visual scratch space." Most multimodal models encode an image once, turn it into tokens, and reason about it in text. Glimpse can return to the source mid-reasoning: zoom into a region, re-watch a short interval at a denser frame rate, or track one feature across time. The chat app shows that scratch space as the model works.

Elorian says a small team built Glimpse on a few hundred GPUs by post-training an open-weights base model, and that it is substantially smaller than the frontier models it compares against. The company did not name the base model in the sources we read.

How Elorian Glimpse scores

The launch table is vendor-run, so read it as a claim. On Elorian's numbers, Glimpse beats GPT-6 Astra on CharXiv-R chart reasoning (92.90 vs 92.10), FSC-147 counting (83.36 vs 82.02), UGC-VideoCap short-video captioning (82.66 vs 80.32), and WGO Bench manipulation events (42.70 vs 39.42), and ties it on ERQA embodied reasoning at 83.75.

It loses elsewhere. Astra leads on BabyVision, scene understanding, dense video captioning, TOMATO temporal reasoning, Video-MME v2, and product counting. Claude Opus 5.5 posts the top CharXiv-R score in the same table. So the honest summary is "competitive and much cheaper," not "best at vision."

The pitch is not that it sees better than everything. It is that it sees well enough for a fraction of the price.

Price is the sharper edge. Elorian says its rates are 94% below GPT-6 Astra, 85% below Opus 5.5, and 70% below GPT-6.1 Sol. In one robot-planning demo with Physical Intelligence's π0.5 policy, Elorian says Glimpse's planning cost about $0.035 for the episode versus $0.249 for Astra. TechCrunch frames it as cheaper than mid-tier models but pricier than the lightest ones.

Why it matters

The target customers are companies drowning in video: robotics firms labeling teleoperation data, labs analyzing charts, media teams tagging clips, and construction or inspection workflows that need accurate counts. Dai told TechCrunch that models still fail at things like reading a board game position or following wordless Ikea diagrams, which humans handle easily.

Elorian's longer-term goal is moving from seeing to doing, including direct robot control. It says calibrated scene reconstruction and action outputs are still future work. For other recent model drops, see Muse Spark Open from Meta and our note on Liquid AI's d1 decision model.

There is no Polymarket market tied to this release, so there is no crowd price to quote here.

The hype check

Jeff Dean's participation and a $300 million seed valuation guarantee attention, and early chatter on X has been warm. But Elorian Glimpse is a preview, the benchmarks are self-reported, and independent leaderboards have not had time to test it. Expect third-party evals over the next few weeks to either confirm the price-performance claim or trim it.

What to do as a reader

If you build with vision, test Elorian Glimpse on your own failure cases: counting with lookalikes, short video clips where order matters, and charts with dense labels. The visual scratch space is the feature to evaluate, because it should show where the model looked, not just what it answered.

Budget for token use, not just token price. Elorian itself says close looking "takes time, and tokens," and lists token efficiency as its next focus. A cheap rate can still produce a big bill if each answer burns many steps. Nothing here is investment advice.

Sources: Elorian launch post and TechCrunch exclusive.