News · SI
Elorian, the Palo Alto startup led by former Google DeepMind researcher Andrew Dai, has released Glimpse as a research preview. The visual reasoning model can zoom into an image, rewatch part of a video or run code while it thinks, and costs $0.60 per million input tokens. Its own benchmark table shows it beating GPT-6 Astra on four tests and trailing it on six.
Elorian has released Glimpse, its first model, as a research preview with API access and a chat interface[1]. The pitch is a model that doesn't look at a picture once and then reason in text: mid-answer, Glimpse can crop into a region, pull specific frames from a video or run code to measure what it's looking at, a mechanism Elorian calls a "visual scratch space"[2].
01 / 04
Elorian describes Glimpse as a visual thinking model that follows objects through motion and changing viewpoints, works out how parts fit together, and connects what it sees to what happens next[1]. It is trained to use tools that write code, transform images and watch videos, and can spawn subagents that pass text and visual artifacts between them[1].
The idea isn't brand new. TechCrunch notes OpenAI added crop-and-zoom during reasoning to its o-series models in 2025; Elorian's argument is that it built the whole model around the approach, cheaply[2]. The company says a small team built Glimpse with a few hundred GPUs, post-training an open-weights base model it doesn't name[1].
02 / 04
Glimpse costs $0.60 per million input tokens and $3.00 per million output tokens, which Elorian puts 94% below GPT-6 Astra's rates and 85% below Claude Opus 5.5's[1]. Elorian is upfront that looking closely takes time and tokens, and says token efficiency is its next priority[1].
On Elorian's own twelve-test table, Glimpse beats Astra on four: manipulation events in robot video (42.70 to 39.42), scientific chart reasoning (92.90 to 92.10), short-form video captioning (82.66 to 80.32) and open-vocabulary counting (83.36 to 82.02)[1]. It ties Astra at 83.75 on embodied spatial reasoning and trails it on six, including Video-MME v2 (56.68 to 64.06) and scene understanding (46.36 to 53.6), while the twelfth test, E.T. Bench, has no Astra score[1]. Claude Opus 5.5 scores highest of all on chart reasoning at 93.40[1].
One demo stands out for robotics teams. Asked to plan for Physical Intelligence's π0.5 robot policy on a two-step task, both Glimpse and Astra completed it, with Glimpse's planning costing $0.035 for the episode against Astra's $0.249[1]. Elorian also labels some of its worked examples as reviewed reference answers rather than recorded model traces, so read the demos with that in mind[1].
03 / 04
Elorian is led by CEO Andrew Dai, a former Google DeepMind researcher, and raised a $55 million seed round at a $300 million valuation before shipping a product[2]. The round was co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with Nvidia and Jeff Dean participating[2]. Its first targets are companies sitting on lots of video: robotics firms, science labs, media and creative businesses[2].
04 / 04
If your product leans on video or images, such as labelling robot data, counting items on site footage or reading charts, Glimpse is a cheap model worth benchmarking against your own data. Vercel has already added it to AI Gateway as elorian/glimpse[3], so trying it is a one-line model swap if you use the AI SDK. Just measure cost per completed task, not per token, because a model that rewatches footage will spend more tokens doing it.
FAQ
Glimpse is Elorian's first model, released on 16 October 2026 as a research preview. Elorian calls it a visual thinking model: it reasons about images and video by returning to the visual evidence mid-answer, cropping into a region, sampling video frames more densely or running code to measure what it sees.
$0.60 per million input tokens and $3.00 per million output tokens. Elorian says both rates are 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. It also says close inspection takes time and tokens, so per-token price isn't the same as per-task cost.
On some tasks. In Elorian's own table it beats Astra on four of twelve evaluations (manipulation events, chart reasoning, short-video captioning and counting), ties it on embodied spatial reasoning, and trails it on six, with the widest gaps on long-video and scene-understanding tests. The table lists no Astra score for the twelfth test, E.T. Bench. No independent group has published results yet.
Through Elorian's API and chat interface. Vercel has also added it to AI Gateway under the model name elorian/glimpse, so it can be called from the AI SDK with image or video inputs.
TechCrunch reports Elorian is a Palo Alto startup led by CEO Andrew Dai, a former Google DeepMind researcher. It raised a $55 million seed round at a $300 million valuation, co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with Nvidia and Google chief scientist Jeff Dean participating.
Not yet. Elorian shows Glimpse planning steps for a separate robot policy and labelling robot footage, but says direct action outputs for robots and calibrated scene reconstruction are still work ahead.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.