Elorian launches Glimpse

On October 16, Palo Alto startup Elorian released its first model, Glimpse. The company, founded by former Google DeepMind researcher Andrew Dai, closed a $55 million seed round at a $300 million valuation roughly six months ago, with Jeff Dean and Nvidia among the investors. Glimpse is not another vision-language model — it is a visual thinking model that repeatedly goes back to the source image mid-reasoning. It can crop into regions, extract specific frames from video, or run code to measure and compare what it is looking at. Elorian calls this mechanism a “visual scratch space.”

Key points

Funding and valuation. Elorian emerged from stealth about six months ago with a $55 million seed round at a $300 million valuation. Striker Venture Partners, Menlo Ventures, and Altimeter Capital co-led, with participation from Nvidia and Google chief scientist Jeff Dean.
Pricing. Glimpse’s API costs $0.60 per million input tokens and $3 per million output tokens — cheaper than Claude Sonnet and GPT-6 Sol, more expensive than Haiku and Luna.
Technical approach. Most multimodal models convert an image to tokens once, then reason about it in text. Glimpse keeps going back to the source mid-reasoning, cropping regions, pulling video frames, or executing code to verify what it sees.

Background

Elorian’s core argument is that today’s AI models cannot “see.” Most vision-language models translate images into tokens, then reason in text space — like describing a photo from memory. Elorian’s homepage states: “Today’s vision language models reason in a two-step process: first translating visual inputs into language, and then performing text-based reasoning. No matter how advanced the subsequent reasoning, this chain is fragile, limited, and prone to hallucination.” CEO Andrew Dai’s motivation came from a game night — he photographed a board game he had just finished and asked a frontier model who was winning. It could not say. Dai told TechCrunch: “Board games each have their own visual language, each have their own iconography, symbols. To an adult, it’s quite easy to understand what’s going on with minimal text, but these models aren’t able to intuit what the symbols mean.” He asked: “Humans can look at IKEA diagrams, with no text, and assemble IKEA furniture, right? Why can’t AI models?”

The facts

Glimpse’s pricing, availability, and self-reported results (Elorian’s official blog and TechCrunch, read October 16):

Item	Number	Whose figure
Seed round	$55 million	TechCrunch, Elorian site
Valuation	$300 million	TechCrunch
Input price	$0.60 / million tokens	Official blog Table 1, TechCrunch
Output price	$3.00 / million tokens	Official blog Table 1, TechCrunch
Availability	API and chat UI both open today	Official blog, models page
Self-reported results	14 evaluations: exceeds GPT-6 Astra on 4, matches on 1, trails on 3	Elorian official blog Table 1
Training scale	Small team with only a few hundred GPUs, post-trained on an open-weights base model	Elorian official blog

On availability, Elorian’s models page has exactly one demo: a seven-second pencil-sketch animation (no audio, no caption) where the model is asked “what kind of dance are the characters doing.” It breaks the action into time segments — 0.0–4.4s “classic closed hold, travelling steadily across the floor,” 4.4–5.4s “he raises his arm and leads an underarm turn — a swing-out,” 5.4–7.3s “open position, facing each other at arm’s length, bouncy side steps.” The page also offers enterprise access (“Bringing Glimpse to a team?”) with volume pricing, custom evaluation, and deployment support, but there is no public pricing table or self-serve signup.

What others say

Elorian’s own HSS benchmark provides the most concrete independent perspective. HSS (Humanity’s Sixth Sense) is a visual reasoning benchmark Elorian built with Scale AI, published October 8. It contains 522 tasks — 288 images and 234 short video clips totaling 17.6 hours. The human baseline scores 93.1%; the best of 25 models, GPT-6 Astra, reaches 53.6%, with a median of 30.9%. Elorian’s blog gives the failure breakdown: across 8,573 labeled failures, 94% come from seeing or inferring rather than reasoning — 53% from misreading something in the image, 41% from failing to infer what the image implied but did not show, and only 5% from faulty logic. HSS also found that giving models agent tools to crop, zoom, and take a second look — essentially what Glimpse does — improved the best score on a 388-task subset from the baseline to 59.3%, cutting failures from 944 to 759.

Beyond the HSS leaderboard there is a fuller set of numbers. Table 1 in Elorian’s official October 16 launch post sets Glimpse against GPT-6 Astra, GPT-6.1 Sol, Opus 5.5 and Gemini 3.8 Flash across 14 evaluations. Glimpse exceeds Astra on four: scientific chart reasoning (CharXiv-R, 92.90 against 92.10), open-vocabulary counting (FSC-147, 83.36 against 82.02), short-form video captioning (UGC-VideoCap, 82.66 against 80.32) and manipulation-event understanding (WGO Bench, 42.70 against 39.42). It matches Astra on embodied spatial reasoning (ERQA, 83.75 each). The official wording is that Glimpse attains “frontier-level performance at the level of (and sometimes exceeding) GPT-6 Astra and Claude Opus 5.5 while being a substantially smaller and cheaper model.”

The same table shows where Glimpse trails. On BabyVision fine-grained visual reasoning it scores 94.33 against Astra’s 97.68; on DREAM-1K dense video captioning 76.04 against 79.05; on TOMATO fine-grained temporal reasoning 93.60 against 96.02. And on HSS itself, Glimpse’s 46.36 falls below Astra’s 53.6 — the top mark in that October 8 leaderboard of 25 models. So “beats GPT-6 Astra” holds only within a defined range: four named evaluations, not visual reasoning in general. Every figure here comes from Elorian’s own blog, with no independent reproduction yet, the same caveat we flagged when covering Meta’s Muse Spark Open — the distance between a vendor’s benchmark and independent reproduction is often more telling than the numbers themselves.

Our take

Glimpse’s story is fundamentally “use HSS’s data to sell Glimpse’s architecture.” Elorian’s own benchmark says clearly: models fail because they misread images — faulty logic accounts for just 5% of failures. Glimpse’s “visual scratch space” is the productized answer to that finding — instead of hoping the model gets it right on the first look, let it re-inspect mid-reasoning. The $0.60/$3 pricing confirms the bet: Glimpse is cheaper than frontier models, which puts Elorian’s competitive edge on inference efficiency rather than model scale.

For developers, the combination of $0.60/$3 pricing and “small team, few hundred GPUs, substantially smaller than frontier systems” raises a real question: if this methodology holds, the cost curve for visual reasoning may be determined by inference architecture, not parameter count. But there is a clear gap in the evidence chain right now — every performance number comes from Elorian itself. The official blog’s 14 evaluations do not say who ran the comparisons or which exact model versions were used, and no independent party has reproduced them. Until someone reruns the suite on their own, treat this as a promising architectural hypothesis, not a verified result.

How to try it

Glimpse’s API is open on the Elorian platform (elorian.ai). The two sources disagree on the chat UI: the official blog says Glimpse “is available through API and our chat UI today,” while TechCrunch reports the chat interface is waitlist-only, so the platform itself is the authority here. If you want to understand what the model does first, Elorian’s models page has the Lindy-Hop dance recognition demo where you can watch the model break down movement into time segments — that is the “re-inspect mid-reasoning” approach in action. The official blog adds three traces with the model’s reasoning visible (a decelerating train, two birds tracked through a camera move, women lifting a tractor tire), showing how it zooms into a region, changes the frame rate on a rewatch, and tracks one feature across time.