Elorian Glimpse is the AI industry’s freshest argument that today’s smartest models still can’t truly see. Released on October 16, 2026, by the Nvidia-backed Palo Alto startup Elorian, Glimpse is a visual-reasoning model built around one simple idea: when the model is unsure about an image, it should look at the image again instead of reasoning from a single, lossy first impression. In this guide, we explain exactly how Glimpse works, what its “visual scratch space” does, how its company-reported benchmarks stack up against GPT-6 Astra, what it costs, and how you can try it right now.
Table of Contents
- What Is Elorian Glimpse?
- Why Today’s AI Can’t Really See
- How Elorian Glimpse Works: The Visual Scratch Space
- Elorian Glimpse Benchmarks: Does It Beat GPT-6 Astra?
- Elorian Glimpse Pricing: 94% Cheaper Than GPT-6 Astra
- Who Is Behind Elorian Glimpse
- How to Try Elorian Glimpse Right Now
- Best Use Cases for Elorian Glimpse
- Elorian Glimpse Limitations: An Honest Look
- Elorian Glimpse vs GPT-6 Astra vs Gemini: Quick Comparison
- The Bottom Line: Should You Care About Elorian Glimpse?
- Elorian Glimpse FAQ
What Is Elorian Glimpse?
Glimpse is the first public AI model from Elorian, a Palo Alto startup founded by former Google DeepMind researcher Andrew Dai. The company came out of stealth in April 2026 with a $55 million seed round at a $300 million valuation — raised on vision alone, before it had shipped any product. Glimpse, released on October 16, 2026, is the first real test of whether that bet was worth it.
Unlike most multimodal models that describe pictures from memory, Glimpse is designed to reason with images. Mid-reasoning, it can zoom into a corner of a chart, pull extra frames out of a video, or even run code to measure what it is looking at. Elorian calls this mechanism a “visual scratch space”, and it is the heart of everything the model does.
Glimpse is launching as a research preview: the API is live now, and a chat interface exists — though reports differ on whether it is waitlist-only or broadly available, which we will untangle below.

Why Today’s AI Can’t Really See
To understand why Glimpse matters, you need to understand the dirty secret of modern multimodal AI. Models like GPT-6 Astra, Gemini, and Claude are brilliant at language, math, and code. But when you hand them a picture, here is what actually happens under the hood:
- They glance once. The model converts the image into tokens — a one-time snapshot.
- They reason about the tokens. From that point on, all “thinking” happens in text, about a compressed representation of the image.
- Details die in translation. Small text in a chart, a subtle symbol in a diagram, the exact position of an object — anything the first glance missed is gone for good.
It is like describing a photograph from memory after glancing at it for one second. You would get the gist right and the details wrong — and that is exactly how frontier models behave today.
Elorian CEO Andrew Dai illustrated the gap with a story from his own game night. Last year, he photographed a board game he had just finished playing with friends and asked a frontier model who was winning. The model couldn’t say. As Dai told TechCrunch, board games each have “their own visual language, their own iconography, symbols” — trivially readable to any adult at the table, but opaque to an AI working from a single tokenized glance. “Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right?” he said. “Why can’t AI models?”
Elorian even helped build a benchmark to quantify the gap. Called Humanity’s Sixth Sense, it quizzes models with 522 questions built around images and video clips. Across those questions, GPT-6 Astra — OpenAI’s current flagship — scored 53.6. Humans averaged 93.1. That chasm between 53.6 and 93.1 is the entire reason Glimpse exists.
How Elorian Glimpse Works: The Visual Scratch Space
The core innovation of Glimpse is disarmingly simple: keep looking. Instead of tokenizing an image once and never returning to it, Glimpse treats the original pixels as a resource it can revisit throughout its reasoning process. Elorian calls this the visual scratch space.
What Elorian Glimpse Can Do Inside the Visual Scratch Space
When Glimpse is working through a visual question and hits uncertainty, it can take several concrete actions that ordinary multimodal models cannot:
- Zoom into a region. Faced with a dense scientific chart, it can crop into one corner and inspect it at higher effective resolution — the AI equivalent of pinching to zoom on your phone photo to double-check what you thought you saw.
- Densify video frames. For video, it can slow down to a denser frame rate over a short interval, catching moments a single sampled frame would miss.
- Track objects across time. It can follow a single object or feature across a sequence of frames mid-reasoning, which is essential for understanding manipulation events and motion.
- Run code to measure. Perhaps most strikingly, the scratch space can execute code to measure or compare what it is observing — turning visual guesswork into computed answers.

How Elorian Glimpse Differs From OpenAI’s Crop-and-Zoom
Sharp-eyed readers will note this is not an entirely new idea. OpenAI added crop-and-zoom during reasoning to its o-series models back in 2025, and an August 2026 research paper called CharTool combined image cropping with code-based calculations. Elorian’s pitch is that it built the whole model around the concept rather than bolting it on. And it did so cheaply: a small team trained Glimpse on only a few hundred GPUs, post-training on an open-weights base model. The result, Elorian says, is a system substantially smaller than the frontier models it benchmarks against — which makes its scores interesting rather than merely promotional.
Elorian Glimpse Is a Native Multimodal Model
Elorian describes its approach as building native multimodal models — systems designed to work directly on images, video, and spatial or physical information without first converting everything into text. Glimpse is, in Elorian’s own words, “a new category of visual thinking model: a system that reasons about the physical world from what it sees.” That framing matters because it points to where the company is headed: from seeing to doing, including direct control of robots.
Elorian Glimpse Benchmarks: Does It Beat GPT-6 Astra?
Elorian’s launch materials include benchmark results that are genuinely eye-catching — with one giant caveat you must understand first: these are company-reported evaluations, not independent tests. As one industry observer put it, it is a test graded by the student who sat the exam. Outside testing will be the real report card.
With that caveat firmly in place, here is what Elorian claims Glimpse achieves against GPT-6 Astra, OpenAI’s current flagship multimodal model:
- Scientific chart reasoning (CharXiv-R): Glimpse 92.90 vs GPT-6 Astra 92.10 — a win on reading and reasoning about research charts.
- Short-form video captioning: Glimpse ahead — describing what happens in short video clips.
- Open-vocabulary object counting: Glimpse ahead — counting arbitrary objects in images, a task that demands actually looking.
- Manipulation-event understanding: Glimpse ahead — understanding physical interactions, directly relevant to robotics.
- Embodied spatial reasoning: tied — understanding space from an agent’s perspective.

Equally important is where Glimpse loses, because Elorian published those numbers too — a refreshing honesty for a launch announcement:
- Scene understanding (HSS benchmark): Glimpse 46.36 vs GPT-6 Astra 53.6 — and the human baseline is 93.1, so both models have a long way to go.
- Dense video captioning and fine-grained temporal reasoning: Glimpse behind.
The honest read: Glimpse is not a frontier model across the board. It is a much smaller, much cheaper model that wins on specific look-closely tasks — which is exactly what its architecture was designed for. The company is upfront about being small, and that candor is part of why the launch has credibility.
Elorian Glimpse Pricing: 94% Cheaper Than GPT-6 Astra
The pricing is where Glimpse stops being an academic curiosity and starts being a business story. Elorian published a price comparison table (Table 1, October 16, 2026) with uncached per-million-token rates:
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Elorian Glimpse | $0.60 | $3.00 |
| GPT-6 Astra | $10.00 | $50.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Gemini 3.8 Flash (intro) | $0.75 | $3.75 |
That puts Glimpse at roughly 94% cheaper than GPT-6 Astra on input tokens — about 17x cheaper on the standard rates. Elorian also gave a concrete robotics example: using Glimpse as the planner for the same robot policy on the same LIBERO task, “Glimpse’s planning costs $0.035 for the episode; Astra’s costs $0.249.”
The strategic logic is clear. Elorian is competing on price per task as much as on benchmark scores. If you run thousands of chart-reading or video-captioning calls per day, a model that is nearly as good (or better) on your specific task at one-seventeenth the price is not a research toy — it is an infrastructure decision. This is the same playbook that made GPT-6 Luna and Claude Haiku 5.5 popular: good-enough intelligence at a fraction of flagship cost.
Who Is Behind Elorian Glimpse
Elorian is a Palo Alto AI startup that, until this week, had more funding than product. The company raised its $55 million seed round at a $300 million valuation in April 2026 — before shipping anything. The round was co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google Chief Scientist Jeff Dean personally. That is an unusually heavyweight cap table for a seed round, and it explains why the AI world paid attention the moment Glimpse dropped.
CEO and co-founder Andrew Dai is a former Google DeepMind researcher. His public thesis is blunt: frontier models handle math, physics, and code reasonably well, but visual understanding has lagged badly behind — and no amount of language-model scaling fixes a model that never looks twice. Elorian’s longer-term ambition goes beyond benchmarks: the company wants to move from seeing to doing, building models that understand a scene well enough to choose actions — including direct robot control. Next year, Elorian says it aims to build “the world’s best visual reasoning model,” one that leads on every visual reasoning test.

How to Try Elorian Glimpse Right Now
Glimpse is a research preview, so access is developer-first for now. Here is the current state of play:
- API access is live. Developers can call the Elorian Glimpse API today. At $0.60 per million input tokens and $3 per million output tokens, it is one of the cheapest vision-capable model APIs on the market.
- Chat interface: reports differ. TechCrunch, which broke the launch story, describes the chat interface as waitlist-only for now — meaning the public launch is, in their words, “a results table and a thesis.” Elorian’s own launch post, however, says Glimpse “is available through API and our chat UI today.” If you want the chat experience, sign up for the waitlist on Elorian’s official site and treat broad availability as pending.
Practical tip: if you are a developer evaluating Glimpse, start with a narrow, look-closely workload — chart extraction, video frame analysis, or visual QA — where its architecture gives it the biggest edge per dollar. Save the general-purpose chat tasks for the bigger frontier models until independent benchmarks confirm broader parity.
Best Use Cases for Elorian Glimpse
Elorian is explicit about where it wants Glimpse to earn its keep: robotics, engineering, medicine, and scientific analysis. Each of these is a domain where getting the visual details right is not cosmetic — it is the whole job. Here is how the model’s strengths map to real work:
- Robotics perception and planning. A robot that misreads its camera feed drops things. Glimpse’s ability to track objects across frames and re-examine regions mid-reasoning maps directly to manipulation tasks — and Elorian’s LIBERO example ($0.035 per planning episode vs $0.249 for Astra) shows the unit economics can work at scale.
- Scientific chart and document analysis. The CharXiv-R win (92.90 vs 92.10) is the headline for a reason: research papers, financial reports, and lab results are full of dense figures where one misread number corrupts the whole analysis. A model that zooms into chart corners before answering is built for this.
- Video understanding at volume. Short-form video captioning at 94%-below-frontier pricing opens up media monitoring, content moderation triage, and video search indexing that would be prohibitively expensive on flagship models.
- Visual quality assurance. Manufacturing inspection, satellite imagery review, medical imaging triage — any pipeline where a human currently squints at images all day is a candidate, with the standard caveat that high-stakes decisions still need human oversight.

Elorian Glimpse Limitations: An Honest Look
No launch-week article would be honest without the other side of the ledger. Here is what tempers the excitement around Glimpse:
- All benchmarks are company-reported. Elorian ran its own evaluations. Independent labs have not re-run them, and history is full of launch-week numbers that softened under outside scrutiny. Treat every win as “claimed” until verified.
- It loses important categories. Glimpse trails GPT-6 Astra on scene understanding, dense video captioning, and fine-grained temporal reasoning. It is a specialist, not a generalist.
- Chat access is limited. If you are not a developer, your path to Glimpse runs through a waitlist. The research-preview framing means rough edges should be expected.
- The thesis is unproven at scale. “Keep looking at the image” is intuitively appealing, and the early numbers support it — but whether the visual scratch space scales to genuinely general visual intelligence is still an open question. Elorian’s own goal — leading every visual reasoning test next year — is an admission that it does not do so today.
- Smaller model, smaller safety surface — but still new. A brand-new model from a brand-new company means less red-teaming history, fewer community evaluations, and thinner documentation than the incumbents. If you are putting Glimpse in production, budget for your own evaluation harness.
Elorian Glimpse vs GPT-6 Astra vs Gemini: Quick Comparison
How does Elorian Glimpse stack up against the models most US readers already know? Here is the honest summary:
| Elorian Glimpse | GPT-6 Astra | Gemini family | |
|---|---|---|---|
| Visual approach | Re-examines images mid-reasoning (visual scratch space) | Single tokenization + text reasoning; crop-and-zoom in o-series lineage | Strong general multimodality; see our Gemini 4 Argon guide |
| Chart reasoning | 92.90 CharXiv-R (claimed) | 92.10 CharXiv-R | Competitive; not the Glimpse comparison target |
| Input price / 1M | $0.60 | $10.00 | $0.75 (Flash intro) |
| Best for | Look-closely visual tasks at scale | General frontier work | Long-context, Google-ecosystem work |
| Maturity | Research preview (Oct 16, 2026) | Flagship, production | Flagship, production |
If your workload is “read ten thousand charts a day,” Elorian Glimpse is suddenly the most interesting model in the room. If your workload is “write my marketing copy and also look at a screenshot sometimes,” the incumbents remain the safer bet.
The Bottom Line: Should You Care About Elorian Glimpse?
Elorian Glimpse is the rare launch that is interesting for two independent reasons. First, the architecture: the visual scratch space is a genuinely different answer to a real, measured problem — frontier vision scores far below human performance, and “look again” is the most human strategy imaginable. Second, the economics: a model that claims targeted wins over GPT-6 Astra at 94% lower prices changes the math for every team running visual AI at volume.
The responsible stance is excitement with a seatbelt: every benchmark is company-reported, the chat product is barely out the door, and the model loses as well as wins. But Elorian did the two things a startup must do on day one — it shipped, and it told the truth about being small. For developers, the API is cheap enough to evaluate this week. For everyone else, Glimpse is worth watching because it points at where multimodal AI is heading: not bigger snapshots, but models that keep looking.
We will update this guide as independent benchmarks land and as Elorian Glimpse moves from research preview to general availability. If you found this useful, you may also like our guides to Google’s free SynthID detector for spotting AI-generated media and the GPT-6 Luna pricing story.
Elorian Glimpse FAQ
What is Elorian Glimpse?
Elorian Glimpse is a visual-reasoning AI model released October 16, 2026, by the Nvidia-backed startup Elorian. It re-examines images and video mid-reasoning — zooming into regions, pulling video frames, and running code to measure what it sees — via a mechanism called the “visual scratch space.”
Is Elorian Glimpse free?
There is no announced free tier. The API is priced at $0.60 per million input tokens and $3 per million output tokens — roughly 94% cheaper than GPT-6 Astra. The chat interface is in limited release (reportedly waitlist-only).
Does Elorian Glimpse beat GPT-6 Astra?
On Elorian’s own evaluations, yes — on scientific chart reasoning, short-form video captioning, open-vocabulary counting, and manipulation-event understanding. It ties on embodied spatial reasoning and loses on scene understanding, dense video captioning, and fine-grained temporal reasoning. These are company-reported numbers, not independent tests.
Who founded Elorian?
Elorian was co-founded by Andrew Dai, a former Google DeepMind researcher. The Palo Alto startup raised $55 million at a $300 million valuation in April 2026, co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with Nvidia and Google Chief Scientist Jeff Dean participating.
How can I try Elorian Glimpse?
The API is live now for developers. A chat interface exists but access reports differ — TechCrunch describes it as waitlist-only, while Elorian’s launch post suggests broader availability.
