Elorian Glimpse is a new visual reasoning model that tries to solve a specific weakness in today's multimodal systems: they look at an image once and then think about it in words. On October 16, 2026, TechCrunch reported that Palo Alto startup Elorian has released Glimpse as a research preview with API access, priced at $0.60 per million input tokens and $3 per million output tokens. Elorian says it beats GPT-6 Astra on two tasks. Those results are the company's own.
This post covers how the model is said to work, the price against models we already track, and a plain accounting of what is claimed versus what is verified.
TL;DR: what we know about Glimpse
| Question | Answer |
|---|---|
| What is it? | A visual reasoning model that can revisit images and video while it reasons |
| Who built it? | Elorian, founded by ex-Google DeepMind researcher Andrew Dai |
| Price | $0.60 per million input tokens, $3 per million output tokens (TechCrunch) |
| Access | Research preview with API access; chat interface waitlist only |
| Headline claim | Beats GPT-6 Astra on scientific chart reasoning and short-form video captioning |
| Independently verified? | Not that we could find |
| Public paper or model card? | None found in our search |
| Funding | $55M seed at a $300M valuation, per TechCrunch; Elorian's site also cites $55M in funding |
How does Glimpse reason differently?
Most multimodal models encode an image into tokens once. After that, the model reasons in text, and anything it failed to capture in that first pass is gone. TechCrunch describes Glimpse as different: mid-reasoning it can return to the source, crop into a region, pull specific frames out of a video, or run code to measure or compare what it sees. Elorian calls this a "visual scratch space."
Bar chart under a magnifying lens, illustrating how Glimpse revisits chart regions while it reasons
The idea resembles what a person does with a diagram: look, form a hypothesis, look again at one corner, measure something with a ruler. It also resembles tool use in coding agents, where the model calls a function instead of guessing. The difference is that the tools here operate on pixels. For a reader who builds with AI, the practical point is that the model's answer can in principle be traced to the crops and measurements it made, rather than to one opaque pass over the image.
Elorian's founder frames the approach as giving visual data equal weight with language tokens. TechCrunch says Elorian came out of stealth about six months before this launch, and that OpenAI added crop-and-zoom reasoning to its o-series models in 2025, so the general approach is not new. Elorian's claim is that it built the capability into the whole model. Elorian's long-term goal, per the same report, is to move from "seeing to doing," including scene models that choose actions and direct robot control.
What does Glimpse cost compared with other models?
TechCrunch puts Glimpse at $0.60 per million input tokens and $3 per million output tokens, and says that sits above lightweight models such as Haiku and Luna and below mid-tier models such as Sonnet and Sol. Those are the lab's tiers as the outlet describes them. We have not seen an Elorian pricing page, so treat the numbers as reported.
Two cautions apply to any visual pricing comparison.
- Image and video tokens are not equal across vendors. A model that crops and re-reads regions may consume more tokens per question than one that looks once. The per-token price tells you little until you measure cost per task on your own images.
- Output tokens include the reasoning. If Glimpse writes out its visual chain of thought, a long analysis of a video could cost more than the headline rate suggests.
For context on how we have compared frontier models on cost, see our GPT-6 Astra launch and pricing breakdown and the Claude Opus 5.5 launch post.
What is claimed and what is verified?
This is the part that matters if you are deciding whether to test Glimpse.
| Claim | Source | Status |
|---|---|---|
| Glimpse can crop, pull frames and run code mid-reasoning | Elorian, via TechCrunch | Described, no public technical write-up found |
| Beats GPT-6 Astra on scientific chart reasoning | Elorian's own evaluations | Unverified; benchmark name and score not given in the report |
| Beats GPT-6 Astra on short-form video captioning | Elorian's own evaluations | Unverified |
| Competitive on other visual reasoning tasks | Elorian's own evaluations | Vague; no numbers reported |
| Trained by a small team on a few hundred GPUs, smaller than rivals | Elorian, via TechCrunch | Unverified; no parameter count disclosed |
| $0.60 in, $3 out per million tokens | TechCrunch | Reported; pricing page not located |
| $55M seed at $300M valuation | TechCrunch | Reported; investors named by TechCrunch |
Notice what is missing: a named benchmark, a score, a model card, a paper, or an independent leaderboard entry. A win on "scientific chart reasoning" could mean a public benchmark or an internal set. Without the name, nobody can reproduce it.
The size claim also deserves care. A small team on a few hundred GPUs training a model that is "substantially smaller" than frontier systems is plausible for a narrow capability, and the broader field has produced several cases of small specialists beating giants on a measured task. But GPT-6 Astra is a general system, and a specialist can beat a generalist on its home turf without being the better choice for your workflow.
How to test Glimpse before you believe it
If you have access, a quick evaluation protects you from both hype and dismissal.
- Build a 50-item set from your own data. Use charts, screenshots, diagrams or video clips where you already know the answer. Include cases where the answer sits in a small region, since that is where a re-inspecting model should shine.
- Run the same set against one model you already use. Compare accuracy and cost per correct answer, not per token.
- Check for visible grounding. If the API returns the crops, frame indices or code it used, verify them. A correct answer with the wrong crop is luck.
- Test video length. Short clips are the claimed strength. Try a clip longer than the demo length before building a pipeline around it.
- Measure latency. Re-inspection loops add round trips. A robotics or live-media use case may care a great deal.
For wider context on how well current models handle video, our guide on whether LLMs can watch video explains the usual workarounds, such as sampling frames. Glimpse's pitch is that the model picks the frames itself.
Where visual reasoning matters for builders
TechCrunch lists three initial markets: robotics firms, scientific labs, and media and creative companies that hold large video libraries. Each maps to a concrete task.
Object detection lens isolating a single object, showing the region-level inspection Glimpse is meant to do
- Robotics: reading a scene, then measuring distances or object relationships in a frame. Our coverage of Gemini 4 Argon for robotics and the robot-arm and humanoid benchmark claims show how fast the frontier labs are moving here, which is the competition Elorian faces.
- Science: reading figures, plots and microscopy images where a single axis label or data point changes the answer.
- Media: captioning, searching and tagging large archives of short video. Captioning is one of the two tasks where Elorian claims a win over GPT-6 Astra.
The pattern across all three is the same: precision on a detail. If your workload is general chat with an occasional photo, a re-inspecting model is probably not worth a new vendor. If a wrong measurement costs money, the idea is worth a pilot.
How does this fit the multimodal landscape?
The frontier labs are all pushing multimodal reasoning. Google's Gemini 4 Argon launch emphasised vision and agent benchmarks, and OpenAI's Astra is the comparison Elorian chose. Academic work also points in the same direction: Apple researchers have published on tri-modal masked diffusion, a very different route to joint text, image and audio models.
What makes Elorian notable is the combination of a seed-stage team and a direct comparison against a flagship. That is a bold framing, and the burden of proof sits with the startup. A reasonable reading is that Glimpse is a credible research preview worth testing, with the benchmark claims treated as marketing until numbers are published.
What to watch next
- A technical report or model card naming benchmarks and scores. This would settle the Astra comparison.
- Independent evaluations from leaderboard operators or researchers.
- Pricing and rate limits on an Elorian page, plus any minimum commitment.
- Chat access moving from waitlist to open.
- Parameter count and context limits, which the report does not give.
Until then, the safest posture is curiosity with a test set.
What this means for what you build or pay
If you build on images or video, add Glimpse to your evaluation shortlist but do not migrate. The reported $0.60 and $3 pricing is moderate, so a pilot is cheap. The decision should rest on cost per correct answer on your own data. If you pay for a general multimodal model today, nothing here changes your bill this week.
Details are as reported on October 16, 2026 and may change as Elorian publishes documentation.