Elorian launches Glimpse to fix AI vision, claiming it beats GPT-6 Astra
Glimpse uses a visual scratch space to re-examine images mid-reasoning, and it costs 94 percent less than GPT-6 Astra.
Models · Source: TechCrunch
What happened
Elorian just released its first AI model called Glimpse. The Palo Alto startup raised a 55 million dollar seed round at a 300 million dollar valuation before launching a single product. Now they are putting that massive valuation to the test. Glimpse is currently available as a research preview with API access and a chat interface.
Founder and former Google DeepMind researcher Andrew Dai says current AI models cannot actually see. Most multimodal models look at an image once, convert it to tokens, and then reason about it purely from memory. That is why they fail at basic visual intuition like reading a board game. Glimpse changes this by keeping the original image accessible. It uses a visual scratch space to crop into regions, pull specific video frames, or run code to measure things mid-reasoning.
The startup claims Glimpse beats OpenAI flagship GPT-6 Astra on scientific chart reasoning and short-form video captioning. A small team trained the entire model using just a few hundred GPUs. It is a specialist model built specifically to solve the visual understanding gap that frontier labs have left wide open. The company built the whole architecture around this crop and zoom reasoning technique.
Key facts
- $55 million — Seed funding raised by Elorian at a $300 million valuation
- $0.60 — Cost per million input tokens for Glimpse
- 94% — How much cheaper Glimpse is compared to GPT-6 Astra
- 24.4% — Average accuracy of AI models on social-reasoning visual tasks
Why it matters
Builders working with robotics, scientific labs, and video media finally have a model built specifically for visual reasoning. You no longer have to rely on a generalist model that loses critical details because it only tokenized an image once. If your pipeline relies on reading dense charts or tracking objects across video frames, you can now query a model that actually zooms in and double-checks its work. This stops the hallucination loop where models confidently guess details they cannot actually see.
This launch proves small teams can still carve out highly profitable niches by ignoring the generalist race. Text benchmarks are completely saturated right now. By focusing entirely on visual reasoning, Elorian built a specialized tool that undercuts flagship model pricing by 94 percent. Developers will start routing specific visual tasks to cheap, narrow specialists rather than paying premium rates to bloated frontier models. The era of one massive model doing everything is fracturing into specialized API calls.
For builders
Route visual tasks to cheaper specialist models
Glimpse costs just 60 cents per million input tokens and 3 dollars per million output tokens. That is 94 percent cheaper than GPT-6 Astra and 85 percent cheaper than Claude Opus 5.5. Startups processing massive video or image datasets can drastically cut compute bills by switching away from generalist APIs.
Build agents that interact with dense visual data
Glimpse can run code to measure or compare what it is looking at mid-reasoning. This unlocks new use cases for agents that need to interpret scientific charts, Ikea diagrams, or complex board games. Companies building visual quality assurance tools or robotics controllers win massive leverage here.
Audit self-reported benchmark claims carefully
Elorian evaluated its own model to claim victory over GPT-6 Astra on charts and video captions. Builders must run their own internal evaluations before ripping out existing OpenAI pipelines. You pay the switching cost if the model fails in production.
My take
I love seeing a small team with a few hundred GPUs take a swing at OpenAI. But remember that Elorian graded its own homework here. Until independent developers test Glimpse against Astra in production, these benchmark wins are just a great marketing pitch.
Original reporting: TechCrunch. This is my rewrite and opinion.