--- title: Elorian releases Glimpse, a vision model that... description: Palo Alto startup Elorian released Glimpse, a research-preview vision model that re-examines images mid-reasoning instead of tokenizing them once. The model... canonical: https://daily.dev/posts/elorian-releases-glimpse-a-vision-model-that-re-examines-images-while-it-reasons-lswi5olu6 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Elorian releases Glimpse, a vision model that re-examines images while it reasons | daily.dev og:description: Palo Alto startup Elorian released Glimpse, a research-preview vision model that re-examines images mid-reasoning instead of tokenizing them once. The model... og:url: https://daily.dev/posts/elorian-releases-glimpse-a-vision-model-that-re-examines-images-while-it-reasons-lswi5olu6 og:image: https://api.daily.dev/og/posts/lswI5OLU6.png og:image:alt: Elorian releases Glimpse, a vision model that re-examines images while it reasons og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Elorian releases Glimpse, a vision model that re-examines images while it reasons **[Collections](https://daily.dev/sources/collections)** · 3 min read · 0 upvotes · 0 comments ## Summary Palo Alto startup Elorian released Glimpse, a research-preview vision model that re-examines images mid-reasoning instead of tokenizing them once. The model maintains a 'visual scratch space,' letting it crop into regions, pull video frames, or run code to re-inspect visual content, aiming to fix failures like misreading board games or IKEA diagrams. Elorian raised $55 million at a $300 million valuation pre-launch, with founders from Google Gemini and Apple MM-1 teams. An informal test showed Glimpse correctly reading 8 of 10 game boards versus Gemini 3.8 Flash's 5 of 10. Elorian claims Glimpse beats GPT-6 Astra on chart reasoning and video captioning despite training on only a few hundred GPUs. Pricing is $0.60 per million input tokens and $3 per million output tokens, positioning it between lightweight and mid-tier models. ## Content Elorian, a Palo Alto startup, has released Glimpse, its first AI model, as a research preview. The company raised $55 million in seed funding at a $300 million valuation before shipping anything. Its team includes people who led work on Google's Gemini and Apple's MM-1, and CEO Andrew Dai is a former Google DeepMind researcher. ## How it differs from other multimodal models Most multimodal models look at an image once, convert it to tokens, and reason about it in text from then on. Glimpse keeps what Elorian calls a "visual scratch space." While it thinks, it can crop into a region of an image, pull a specific frame from a video, replay half a second at a higher frame rate, or run code to look at the visual content again before answering. Dai says this targets failures like models being unable to read board-game states or IKEA diagrams. ## Reported results The numbers below come from Elorian's own tests and a few hand-checked demos, not independent benchmarks. - Elorian says Glimpse beats GPT-6 Astra on scientific chart reading and short-form video captioning, despite being trained on only a few hundred GPUs. - In a robot-arm simulation, planning a task where the arm places cream cheese in a bowl cost $0.035 with Glimpse versus $0.249 with GPT-6 Astra. - In a maze demo, Glimpse solved 5 mazes of increasing difficulty. Gemini 3.8 Flash solved 1 of 5. - @omarsar0 ran 10 game boards photographed from overhead with the same prompt and image on both models. Glimpse got 8 of 10 fully correct, and Gemini 3.8 Flash got 5 of 10. Pricing is $0.60 per million input tokens and $3 per million output tokens, which puts it between lightweight and mid-tier models. ## Why people care The reaction has centered on video and robotics. @kloss_xyz compared the usual approach to doing a 500-piece jigsaw after glancing at the box for two seconds and never looking again. Glimpse keeps looking: it zooms in, pulls the exact frame, and rewatches a moment it's unsure about. @kloss_xyz listed robots, job sites, factory lines, game replays, and even counting animals at home as possible uses. @svpino pointed out that it can describe how objects in an image relate to each other, or how an action changes over time in a video, and said robotics teams would want that. @trikcode framed the problem as models that describe one photo well and then lose the thread when the camera moves. Glimpse is meant to follow things across frames and angles. @omarsar0 made the data-labeling point. Getting a whole frame right, not just most of it, is what turns raw robot footage into usable training data. I'd treat the comparisons cautiously until outside groups test it. Still, a model that can go back and look again is a sensible idea, and the maze and board results suggest it does something real. ## Questions this post answers ### What makes Elorian's Glimpse model different from other vision-language models? Glimpse maintains a visual scratch space during reasoning, letting it crop into image regions, pull specific video frames, or run code to re-examine visual content multiple times rather than tokenizing an image once and reasoning purely in text afterward. This targets failures like misreading board game states or following IKEA diagrams, and tracking objects across video frames and camera angles. _Teams comparing vision models for robotics or video data pipelines track releases like this on daily.dev._ ### How much does Elorian's Glimpse vision model cost to use via API? Glimpse is priced at $0.60 per million input tokens and $3 per million output tokens, placing it between lightweight and mid-tier multimodal models. It was released as a research preview by Elorian, a Palo Alto startup that raised $55 million in seed funding at a $300 million valuation before shipping any product. _Developers weighing model costs against accuracy for vision tasks follow pricing comparisons like this on daily.dev._ ### How did Glimpse perform against Gemini 3.8 Flash on reading game board images? In a hand-checked test of 10 game boards photographed from an overhead camera with identical prompts, Glimpse correctly read 8 of 10 boards while Gemini 3.8 Flash got 5 of 10. The sample size is small, but the task is relevant to turning raw robot footage into usable training data, where getting the full frame right matters. _Engineers validating vision models before adoption keep an eye on early benchmark reports via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 27 comments across x (as of 2026-10-17). **TL;DR:** Reactions are curious and cautiously impressed by the idea of a model that rewinds/zooms into video mid-reasoning, but several commenters push back that the benchmarks are self-reported and not a clean win over rivals. **Sentiment:** 30% positive · 45% mixed · 25% skeptical **The case for** - The ability to rewatch or zoom into specific moments to resolve uncertainty is seen as a genuinely useful, human-like approach. - The pricing gap (cited as ~7x cheaper on a robot-arm planning task) is seen as compelling even if the model isn't always the top performer. - Temporal/spatial reasoning improvements are seen as especially relevant for robotics use cases. **The pushback** - Benchmarks are self-designed (own mazes, own agent comparisons), so winning against competitors on them is viewed with skepticism. - On the company's own benchmark it reportedly still trails a competitor (46.4 vs 53.6), undercutting claims of a clean sweep. - Unclear whether the cost advantage holds once extra frames from replays/higher frame rate are factored into token costs. - No independent evaluation has been shown yet, and bold claims like 'goes way beyond frontier models' lack named benchmarks. **By community** - x (mixed): Replies mix genuine interest in the rewind/zoom mechanism with pointed questions about cherry-picked demos, self-reported benchmarks, and real token costs. **Hottest debate:** Whether the touted benchmark wins and cost savings are meaningful given they come from Elorian's own self-designed tests rather than independent evaluation. **Open questions** - Is the model open weights or API only? - Does per-token API pricing account for the extra frames pulled in during replays? - Does it still perform well when camera angle changes mid-video? - Does it decide on its own when to zoom in, or does it need to be prompted? **Highlights** > @testingcatalog The cost number is the real story here. $0.035 vs $0.249 for the robot arm plan is roughly 7x cheaper. Funny part is on their own Sixth Sense benchmark it still trails Astra, 46.4 vs 53.6, so it's not a clean sweep. Cheap and good enough usually wins anyway. > — [RealGesar on x · 1 comments](https://x.com/RealGesar/status/2111175766041931865) > @svpino "Goes way beyond frontier models" on spatial and temporal reasoning is a big claim with zero benchmark named. Show it losing to GPT-4V or Gemini on something like CLEVRER or EgoSchema, not just a cherry-picked clip. > — [ukrroot on x](https://x.com/ukrroot/status/2111161798305210546) > @testingcatalog Adaptive visual reinspection can reduce wasted inference, but benchmark gains matter only when they translate to reliable real-world task completion > — [SlimAssiliX on x · 1 points](https://x.com/SlimAssiliX/status/2111179879743983755) > @testingcatalog its own tests, its own mazes, its own agent against gemini flash. anybody can win a race they designed > — [rayvahey on x · 1 points](https://x.com/rayvahey/status/2111175889371517086) > @testingcatalog $0.035 vs $0.249 works out to about 7x cheaper on that task. does the cost count the extra frames it pulls in when it replays at a higher frame rate? on longer videos those re-watches seem like they'd add input tokens fast > — [Chahatusharma on x](https://x.com/Chahatusharma/status/2111179916159099212) **Source threads** - [x](https://x.com/kloss_xyz/status/2111147758774075556) · 0 points · 4 comments - [x](https://x.com/svpino/status/2111150762944516216) · 0 points · 9 comments - [x](https://x.com/testingcatalog/status/2111174690337177607) · 1 points · 14 comments ## Similar posts on daily.dev - [How Glance turns hours of video into mobile-ready clips with AI](https://daily.dev/posts/how-glance-turns-hours-of-video-into-mobile-ready-clips-with-ai-vikjn8jaa) · Google Cloud · 1 upvotes · 0 comments --- Tags: [#computer-vision](https://daily.dev/tags/computer-vision), [#google-gemini](https://daily.dev/tags/google-gemini), [#multimodal](https://daily.dev/tags/multimodal) [View this post on daily.dev](https://daily.dev/posts/elorian-releases-glimpse-a-vision-model-that-re-examines-images-while-it-reasons-lswi5olu6) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Elorian releases Glimpse, a vision model that re-examines images while it reasons","url":"https://daily.dev/posts/elorian-releases-glimpse-a-vision-model-that-re-examines-images-while-it-reasons-lswi5olu6","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/elorian-releases-glimpse-a-vision-model-that-re-examines-images-while-it-reasons-lswi5olu6"},"datePublished":"2026-10-16T17:23:23.516Z","dateModified":"2026-10-17T07:49:18.446Z","description":"Palo Alto startup Elorian released Glimpse, a research-preview vision model that re-examines images mid-reasoning instead of tokenizing them once. The model...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/elorian-releases-glimpse-a-vision-model-that-re-examines-images-while-it-reasons-lswi5olu6","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"computer-vision,google-gemini,multimodal","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Elorian releases Glimpse, a vision model that re-examines images while it reasons"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/elorian-releases-glimpse-a-vision-model-that-re-examines-images-while-it-reasons-lswi5olu6#faq","mainEntity":[{"@type":"Question","name":"What makes Elorian's Glimpse model different from other vision-language models?","acceptedAnswer":{"@type":"Answer","text":"Glimpse maintains a visual scratch space during reasoning, letting it crop into image regions, pull specific video frames, or run code to re-examine visual content multiple times rather than tokenizing an image once and reasoning purely in text afterward. This targets failures like misreading board game states or following IKEA diagrams, and tracking objects across video frames and camera angles. Teams comparing vision models for robotics or video data pipelines track releases like this on daily.dev."}},{"@type":"Question","name":"How much does Elorian's Glimpse vision model cost to use via API?","acceptedAnswer":{"@type":"Answer","text":"Glimpse is priced at $0.60 per million input tokens and $3 per million output tokens, placing it between lightweight and mid-tier multimodal models. It was released as a research preview by Elorian, a Palo Alto startup that raised $55 million in seed funding at a $300 million valuation before shipping any product. Developers weighing model costs against accuracy for vision tasks follow pricing comparisons like this on daily.dev."}},{"@type":"Question","name":"How did Glimpse perform against Gemini 3.8 Flash on reading game board images?","acceptedAnswer":{"@type":"Answer","text":"In a hand-checked test of 10 game boards photographed from an overhead camera with identical prompts, Glimpse correctly read 8 of 10 boards while Gemini 3.8 Flash got 5 of 10. The sample size is small, but the task is relevant to turning raw robot footage into usable training data, where getting the full frame right matters. Engineers validating vision models before adoption keep an eye on early benchmark reports via daily.dev."}}]} ```