Glimpse (Elorian)
Glimpse is a multimodal AI model from Elorian, a Palo Alto startup led by former Google DeepMind researcher Andrew Dai. Elorian released it as a research preview on 16 October 2026 and calls it a "visual thinking model": a system meant to reason about the physical world from images and video, returning to the visual evidence while it works instead of turning a picture into text once and reasoning over the text.[1][3] It is Elorian's first model.[2][3] Elorian says Glimpse is post-trained on an open-weights base model, which it does not name, and that a small team developed it "with only a few hundred GPUs".[1] The model is sold through an API and a chat interface at $0.60 per million input tokens and $3.00 per million output tokens.[1] In a benchmark table published by Elorian, Glimpse scores above or level with GPT-6 Astra on five of the eleven evaluations where Astra has a score and below it on the other six.[1]
Background
Elorian argues that current vision language models work in two steps: they first translate visual input into language and then reason in text. The company's website calls that chain "fragile, limited, and prone to hallucination" and says many visual tasks "cannot be precisely (or concisely) described" in words.[4] Andrew Dai, Elorian's co-founder and chief executive, described the problem in his launch post: "Spent years watching models ace a single image and completely whiff on anything that moves or changes angle."[2] He told TechCrunch that a frontier model he tested could not say who was winning in a photo of a board game, and asked, "Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can't AI models?"[3]
Dai previously worked at Google Brain and DeepMind, where Elorian's team page credits him with co-leading pretraining for PaLM 2 and leading the Gemini data area. Co-founder Yinfei Yang previously led multimodal foundation-model work at Apple, including the MM1 and MM1.5 model families.[5] Elorian came out of stealth about six months before Glimpse's release, with a $55 million seed round that TechCrunch reported valued the company at $300 million.[3][8] The company's website lists Striker Ventures, Menlo Ventures and Altimeter as investors, with participation from NVIDIA, 49 Palms and individual researchers including Jeff Dean.[4] On 8 October 2026, Elorian and Scale AI published Humanity's Sixth Sense (HSS), a benchmark of open-ended visual reasoning questions that the company also uses to evaluate Glimpse.[6]
Design
Elorian's launch post does not name the base model or give a parameter count, and the company's description of the model's design comes from that post.[1] According to Elorian, Glimpse:
- can "reason in visual space, revisiting and refining internal visual representations as thinking unfolds", and learns to reason about objects and their spatial relationships;[1]
- is trained end-to-end, as an agent, to use multimodal tools that write code, transform images and watch videos;[1]
- can spawn and coordinate subagents that exchange text, visual artifacts and internal representations to work on a shared problem.[1]
In the chat interface the platform displays what Elorian calls the model's "visual scratch space" while it reasons. Elorian says the model will zoom into a region, watch a short interval at a denser frame rate, or track one feature across time.[1] TechCrunch described the same mechanism as cropping into a region, pulling specific frames from a video, or running code to measure what the model is looking at in the middle of its reasoning, and noted that the idea is not entirely new: OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. Elorian's pitch, TechCrunch wrote, is that it built the whole model around the approach and did so cheaply.[3] Elorian describes Glimpse as "substantially smaller and cheaper" than the frontier models it compares against, without giving a size.[1]
Availability and pricing
Elorian says Glimpse is available through an API and an interactive chat interface on the Elorian platform, and the launch post invites users to "try Glimpse free".[1][7] TechCrunch's launch-day report described the chat interface as waitlist-only.[3] Elorian's models page invites companies to contact its sales team for volume access, custom evaluation against their own data and deployment support.[9]
Elorian published the following prices alongside its benchmark table (US dollars per million tokens; input prices are uncached).[1]
| Model | Input | Output | Glimpse's price as a share (computed) |
|---|---|---|---|
| Glimpse | $0.60 | $3.00 | 100% |
| GPT-6 Astra | $10.00 | $50.00 | 6% |
| Claude Opus 5.5 | $4.00 | $20.00 | 15% |
| GPT-6.1 Sol | $2.00 | $10.00 | 30% |
| Gemini 3.8 Flash (introductory) | $0.75 | $3.75 | 80% |
Elorian's stated discounts follow directly from these figures: Glimpse's input and output rates are each 94% lower than GPT-6 Astra's, 85% lower than Claude Opus 5.5's, 70% lower than GPT-6.1 Sol's and 20% lower than Gemini 3.8 Flash's introductory rates.[1] The post's social-preview description expresses the Astra comparison as token prices 16.7 times lower.[1] TechCrunch placed Glimpse's price above lightweight models such as Haiku and Luna and below mid-level models such as Sonnet and Sol.[3]
Benchmark results
All of the results below are Elorian's own figures from its launch table.[1] Elorian ran each model at its highest available thinking effort with what it calls the model's native agent harness: Glimpse at max; GPT-6 Astra, GPT-6.1 Sol and Opus 5.5 at max with Code Interpreter; and Gemini 3.8 Flash at high with Code Interpreter for images and with Video Agent for video. For video evaluations, GPT-6.1 Sol and Opus 5.5 were run at high effort with Code Interpreter, except Sol's Video-MME v2 score, which uses max; Elorian says max-effort results were not available for the other video evaluations. A dash means no result was reported.[1]
| Capability | Evaluation | Glimpse | GPT-6 Astra | GPT-6.1 Sol | Opus 5.5 | Gemini 3.8 Flash |
|---|---|---|---|---|---|---|
| Robotics | WGO Bench (manipulation events) | 42.70 | 39.42 | 38.91 | 12.34 | 33.99 |
| Robotics | ERQA (accuracy) | 83.75 | 83.75 | 82.45 | 82.20 | 81.20 |
| Visual perception | BabyVision | 95.36 | 97.68 | 97.16 | - | 83.25 |
| Visual perception | HSS (image and video) | 46.36 | 53.6 | 46.6 | 49.81 | 41.6 |
| Scientific reasoning | CharXiv-R | 92.90 | 92.10 | 92.60 | 93.40 | 91.80 |
| Video captioning | UGC-VideoCap | 82.66 | 80.32 | 80.41 | 76.35 | 78.33 |
| Video captioning | DREAM-1K | 76.04 | 79.05 | 77.93 | 72.05 | 70.00 |
| Temporal reasoning | TOMATO | 93.60 | 96.02 | 89.02 | 72.44 | 83.02 |
| Temporal reasoning | Video-MME v2 with subtitles (grouped score) | 56.68 | 64.06 | 65.27 | - | 59.58 |
| Temporal reasoning | E.T. Bench (event understanding) | 63.55 | - | 59.79 | - | 63.45 |
| Counting | FSC-147 | 83.36 | 82.02 | 82.69 | 73.78 | 67.06 |
| Counting | CountQA | 82.66 | 86.42 | 86.42 | 82.20 | 81.09 |
Bold marks the highest score in each row. Elorian summarized the table as Glimpse exceeding Astra on four evaluations (scientific chart reasoning, open-vocabulary counting, short-form video captioning and manipulation-event understanding) and matching it on embodied reasoning.[1] Read row by row:
- Where Glimpse leads. It has the top score on WGO Bench, UGC-VideoCap and FSC-147, ties Astra on ERQA, and narrowly leads Gemini 3.8 Flash on E.T. Bench (63.55 against 63.45), where no Astra score was reported. Its largest margin is on WGO Bench, where Claude Opus 5.5 scores 12.34.[1]
- Where it is close. On CharXiv-R it is 0.8 points above Astra but 0.5 points below Claude Opus 5.5, which has the best score in that row.[1]
- Where it trails. It is below Astra on BabyVision (by 2.32 points), HSS (7.24), DREAM-1K (3.01), TOMATO (2.42) and CountQA (3.76). On Video-MME v2 it has the lowest score of the four models reported, 8.59 points behind GPT-6.1 Sol and below Gemini 3.8 Flash.[1]
Elorian notes that its HSS row "includes paper-reported results".[1] The GPT-6 Astra (53.6), GPT-6.1 Sol (46.6) and Gemini 3.8 Flash (41.6) entries match the HSS paper's leaderboard as reproduced on Elorian's HSS post. That leaderboard lists Claude Opus 5.5 at 44.6 at "xhigh" effort, while the Glimpse table lists 49.81; Elorian does not explain the difference, though its table says Opus 5.5 was run at max effort.[1][6] Glimpse itself does not appear on the HSS leaderboard in that post, whose human baseline is 93.1.[6]
Evaluations used
The table mixes published academic benchmarks, a small dataset hosted on Hugging Face, and HSS, which Elorian co-created. The descriptions below are from each benchmark's own paper or dataset card.
| Benchmark | What it tests | Source |
|---|---|---|
| WGO-Bench ("What's Going On Benchmark") | A small, manually annotated set of 100 robot and egocentric manipulation videos with 743 gold subtask segments; models detect where one manipulation subtask ends and the next begins, and describe each completed manipulation event. Elorian's table calls its row "WGO Bench" without linking a source. | [10] |
| ERQA (Embodied Reasoning Question Answering) | 400 multiple-choice visual questions on spatial, trajectory and action reasoning, state estimation, pointing, multi-view and task reasoning, introduced with Gemini Robotics. | [11] |
| BabyVision | 388 items in 22 subclasses testing core visual abilities "independent of linguistic knowledge"; the paper reports an average adult score of 94.1. | [12] |
| HSS (Humanity's Sixth Sense) | 522 open-ended tasks (288 images and 234 short video clips) on what a scene implies but does not show, graded against human-written rubrics; built by Scale AI and Elorian. | [6] |
| CharXiv (reasoning questions) | 2,323 charts from arXiv papers; the reasoning subset asks questions that require synthesizing information across complex chart elements. | [13] |
| UGC-VideoCap | Detailed audio-visual captioning of 1,000 short-form TikTok videos, with 4,000 question-answer pairs. | [14] |
| DREAM-1K | A video-description benchmark with videos from diverse sources and an automatic method for scoring fine-grained descriptions, introduced with the Tarsier models. | [15] |
| TOMATO | 1,484 human-annotated questions on 1,417 videos covering action count, direction, rotation, shape and trend, velocity and frequency, and visual cues; built around principles such as multi-frame gain and frame-order sensitivity, because the authors found many questions in earlier benchmarks could be answered from a single, few or out-of-order frames. | [16] |
| Video-MME-v2 | A video-understanding benchmark that scores groups of related questions non-linearly, giving credit only for consistent, coherent answers rather than isolated correct guesses. | [17] |
| E.T. Bench | 7,300 samples across 12 event-level, time-sensitive video tasks drawn from 7,000 videos. | [18] |
| FSC-147 | A counting dataset of over 6,000 images across 147 object categories, originally built for few-shot counting from exemplar objects. | [19] |
| CountQA | Over 1,500 question-answer pairs on real-world images with high object density, clutter and occlusion. | [20] |
Demonstrations
Elorian's launch post includes several worked examples and labels which ones are recorded model output and which are reference answers written by people.[1]
| Example | What Elorian shows | Type, per Elorian |
|---|---|---|
| Slowing train | Glimpse tracks a window edge from about 2.2 seconds to 5.13 seconds and concludes the train is decelerating because it travels less distance per interval. | Recorded model output, trace condensed |
| Two birds and a camera move | Glimpse captions a 6.13-second clip, inspects it at 8 fps, then rewatches a 0.9-second interval at about 14 fps to decide whether a transition is a continuous pan or a hard cut. | Recorded evaluation run, caption condensed |
| Lifting a tractor tire | Predicts which person is at risk if one loses her grip. | Reference answer, not a model trace |
| Railway switch | Explains how moving a lever changes the track points. | Reference answer, not a model trace |
| Counting open pliers | Separates open pliers from closed pliers and scissors; reference count is four. | Reference answer, not a model trace |
The post also shows applied demonstrations. Glimpse labels and boxes objects in teleoperation footage and splits one episode into six timestamped pick-and-place events. In a planning demo on a LIBERO task, Glimpse and GPT-6 Astra each acted as the high-level planner for Physical Intelligence's π0.5 vision-language-action policy; Elorian says both completed the task, with planning costing $0.035 for the episode with Glimpse and $0.249 with Astra.[1] Other clips cover dense captioning of a street performance in Havana and a 16-second stock clip, where Elorian says Glimpse caught details that Gemini 3.8 Flash and GPT-6.1 Sol missed; counting puppies in a handheld phone video and construction vehicles on a job site; mazes and spot-the-difference puzzles; and twelve randomly drawn HSS "what happens next" questions. In the HSS clip, the motion shown after each answer was generated from the still image by Seedance 2.5 to illustrate the reference answer, and no model saw it.[1] A separate example on Elorian's models page shows Glimpse identifying a dance in a seven-second pencil sketch as a swing-style partnered dance.[9]
Limitations and plans
Elorian describes Glimpse as "a preview" and "an early step". It says calibrated reconstruction and robot control "remain work ahead", and that direct action outputs for robots are a next step it is developing; the railway-switch example notes that direct action generation is future work.[1] The company says it is working toward multimodal-native pretraining and models that both understand and generate.[1] TechCrunch reported that Elorian's initial targets are companies with large amounts of video, such as robotics firms, scientific labs, and media and creative companies, and that Dai said the company aims next year to build "the world's best visual reasoning model".[3]
Glimpse's benchmark comparisons are self-reported, use different effort settings for some competitors on video tasks, and include one benchmark (HSS) co-developed by Elorian.[1][6]
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28Elorian. "Introducing Glimpse." Elorian blog, 16 October 2026. elorian.ai/...introducing-glimpse
- ^1 ^2Andrew M. Dai. Post on X announcing Glimpse, 16 October 2026. x.com/...2111142161584628060
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Aditya Mehta. "Elorian's first AI model wants to reason with images, not just about them." TechCrunch, 16 October 2026. techcrunch.com/...-with-images-not-just-about-them
- ^1 ^2Elorian. Company homepage. elorian.ai
- ^Elorian. "Our Team." elorian.ai/team
- ^1 ^2 ^3 ^4 ^5Elorian. "Introducing Humanity's Sixth Sense (HSS)." Elorian blog, 8 October 2026. elorian.ai/...humanitys-sixth-sense
- ^Elorian. "Support." elorian.ai/support
- ^Maggie Nye and Isabelle Johannessen. "How Elorian AI pulled off a $300M pre-seed valuation." TechCrunch Build Mode podcast, 16 July 2026. techcrunch.com/...ed-off-a-300m-pre-seed-valuation
- ^1 ^2Elorian. "Models." elorian.ai/models
- ^Macrodata. "WGO-Bench: What's Going On Benchmark." Hugging Face dataset card. huggingface.co/...WGO-Bench
- ^Gemini Robotics Team. "Gemini Robotics: Bringing AI into the Physical World." arXiv:2503.20020, 2025. arxiv.org/...2503.20020
- ^Liang Chen et al. "BabyVision: Visual Reasoning Beyond Language." arXiv:2601.06521, 2026. arxiv.org/...2601.06521
- ^Zirui Wang et al. "CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs." arXiv:2406.18521, 2024. arxiv.org/...2406.18521
- ^Peiran Wu et al. "UGC-VideoCaptioner: An Omni UGC Video Detail Caption Model and New Benchmarks." arXiv:2507.11336, 2025. arxiv.org/...2507.11336
- ^Jiawei Wang et al. "Tarsier: Recipes for Training and Evaluating Large Video Description Models." arXiv:2407.00634, 2024. arxiv.org/...2407.00634
- ^Ziyao Shangguan et al. "TOMATO: Assessing Visual Temporal Reasoning Capabilities in Multimodal Foundation Models." arXiv:2410.23266, 2024. arxiv.org/...2410.23266
- ^Chaoyou Fu et al. "Video-MME-v2: Towards the Next Stage in Benchmarks for Comprehensive Video Understanding." arXiv:2604.05015, 2026. arxiv.org/...2604.05015
- ^Ye Liu et al. "E.T. Bench: Towards Open-Ended Event-Level Video-Language Understanding." arXiv:2409.18111, 2024. arxiv.org/...2409.18111
- ^Viresh Ranjan et al. "Learning To Count Everything." arXiv:2104.08391, 2021. arxiv.org/...2104.08391
- ^Jayant Sravan Tamarapalli et al. "CountQA: How Well Do MLLMs Count in the Wild?" arXiv:2508.06585, 2025. arxiv.org/...2508.06585
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
2 revisions · v2 · 2,621 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent verification 2026-10-17 (xg19 V1): 34 sources, ~210 claims; 60 benchmark cells and price arithmetic checked against Elorian's post; 3 minor fixes applied.
Cite this page: AI Wiki. "Glimpse (Elorian)." aiwiki.ai, updated 16 Oct 2026, fact-checked 16 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/glimpse_elorian