Why it matters
Most chatbots look at a picture once and then reason in words. Glimpse is built to go back and look again, at a crop of an image or a split second of video, before it answers. Here is what Elorian showed, what other labs did first, and what nobody outside has checked yet.
I think seeing is where the next race is leaning towards. Reading and writing is solved. Training their eyes for stills and film is still in progress.
What the model looks at
What the model looks at
Elorian's platform shows what it calls the model's visual scratch space. Glimpse can zoom into a region, watch a short interval at a denser frame rate or track a feature across time.
In the recorded birds run, it watches the full 6.13-second clip at 8 frames per second, then revisits 4.3 to 5.2 seconds at near 14 frames per second to distinguish a camera pan from a cut.
In Elorian's construction-count session, Glimpse answers about 30 vehicles and lists the timestamps it relied on.
Sources: Elorian: Introducing Glimpse and Elorian: Construction-count session
What Elorian's benchmark table says
What Elorian's benchmark table says
On Elorian's own table, Glimpse scores above GPT-6 Astra on four evaluations: WGO Bench, CharXiv-R, UGC-VideoCap and FSC-147. It ties on ERQA.
It scores below GPT-6 Astra on six: BabyVision, HSS, DREAM-1K, TOMATO, Video-MME v2 and CountQA. These are Elorian's reported scores. Our search found no independent test of Glimpse as of 29 October 2026.
Token prices and demo conditions
Token prices and demo conditions
Glimpse is priced at USD 0.60 per million input tokens and USD 3.00 per million output tokens. Elorian says both rates are 94 percent below GPT-6 Astra's.
Elorian says looking closely takes time and tokens, and names token efficiency as its next focus.
Its robot-planner demo reports USD 0.035 for Glimpse and USD 0.249 for Astra for one episode. The effort settings differ: Glimpse runs at default effort, Astra at high.
The maze demo reports Glimpse at 20/20, GPT-6 Sol at 10/20 and Gemini 3.8 Flash at 3/18. Glimpse had tools enabled; the others had none.
| Model | Input, USD per million tokens | Output, USD per million tokens |
|---|---|---|
| Elorian Glimpse | 0.60 | 3.00 |
| GPT-6 Astra | 10.00 | 50.00 |
| GPT-6.1 Sol | 2.00 | 10.00 |
| Claude Opus 5.5 | 4.00 | 20.00 |
| Gemini 3.8 Flash | 0.75 | 3.75 |
From the price table in Elorian's launch post, 16 October 2026. Gemini 3.8 Flash at its introductory price.
Sources: Elorian: Introducing Glimpse, pricing and robot-planner demo and Elorian: Visual puzzles demo and settings
Who built it
Who built it
Elorian's co-founders include CEO Andrew Dai, formerly of Google Brain and DeepMind and a Gemini data area lead, and Yinfei Yang, formerly of Apple, where he led the MM1 and MM1.5 multimodal models.
Elorian says a small team built Glimpse on a few hundred GPUs and post-trained it on an unnamed open-weights base model.
Sources: Elorian: Team and Elorian: Introducing Glimpse, training details
The second look has a history
The second look has a history
- 25 September 2023 OpenAI rolls out image input in ChatGPT. Elorian and TechCrunch describe the usual way since: the picture becomes language or tokens once, then the model reasons in text.
- 16 April 2025 OpenAI's o3 and o4-mini crop, zoom and rotate images within their reasoning. As TechCrunch notes, crop-and-zoom reasoning predates Glimpse.
- 27 January 2026 Google's Agentic Vision in Gemini 3 Flash writes and runs Python to crop, rotate or annotate an image, then feeds the transformed image back into context.
- 1 September 2026 Google's agentic video in Gemini moves beyond a default of 1 frame per second: the model decides what to watch and can revisit intervals at a higher frame rate.
- 16 October 2026 Elorian releases Glimpse, a model built around that second look.
Sources: OpenAI: ChatGPT can now see, hear, and speak, Elorian: Model overview, TechCrunch: Elorian's first AI model, OpenAI: Thinking with images, Google: Agentic Vision in Gemini 3 Flash, and Google: Introducing agentic video in Gemini
Read the sources
Read the sources
Every claim above comes from these pages. The scores are Elorian's own tests, so the date we read them is next to each.
- Read the postElorian: Introducing Glimpse
The launch post: how Glimpse looks again, the worked examples, the benchmark table and the prices.
Our screenshot of Elorian's blog, 29 October 2026 - Read Google's postGoogle: Agentic Vision in Gemini 3 Flash
January 2026: Gemini writes code to crop, rotate or mark up an image while it reasons.
Our screenshot of Google's blog, 29 October 2026 - Read Google's postGoogle: agentic video in Gemini
September 2026: Gemini decides what to rewatch and at what frame rate.
- Read OpenAI's postOpenAI: Thinking with images
April 2025: o3 and o4-mini crop and zoom into images inside their reasoning.
- Read the storyTechCrunch: Elorian's first AI model
The launch story, with how Elorian came out of stealth.
No independent test of Glimpse was found as of 29 October 2026. Every score on this page is Elorian's own run.
Yes, that's my AI avatar in the video. Same opinions, better lighting.