Palo Alto startup Elorian has unveiled Glimpse, its debut AI model and a direct challenge to the assumption that multimodal systems should convert a picture into tokens once and then reason about it from memory. Released on 16 October as a research preview, the model ships with strikingly aggressive pricing — $0.60 per million input tokens and $3.00 per million output tokens — and a benchmark table that puts it alongside, and on a handful of tasks ahead of, far larger frontier systems.
- A visual scratch space: Glimpse can crop into a region, zoom, pull denser video frames and execute code mid-reasoning, returning to internal visual representations instead of relying on a single text description.
- Price shock: Elorian claims its rates sit 94% below GPT-6 Astra, 85% below Claude Opus 5.5, 70% below GPT-6.1 Sol and 20% below Gemini 3.8 Flash’s introductory pricing.
- A split scoreboard: Glimpse beats GPT-6 Astra on four evaluations and ties on one, but trails on six — all measured by Elorian itself, with no independent benchmarking yet published.
- A small-team build: A few hundred GPUs and post-training on an unnamed open-weights base model, with no parameter count disclosed.
Reasoning inside visual space, not merely about it
Elorian describes Glimpse as a new category of “visual thinking” model. Rather than inspecting an image once, it revisits and refines internal visual representations as its reasoning unfolds, tracking objects through motion and shifting viewpoints, parsing how parts fit together, and linking an observation to what happens next. The company calls this loop a “visual scratch space”, and says the platform shows the model zooming into a region, watching a short interval at a denser frame rate, or running code to measure what it is studying.
When deployed as an agent, Glimpse is trained end-to-end to use multimodal tools that write code, transform images and watch video. It can also spawn and coordinate subagents that exchange text, visual artifacts and internal representations to solve a shared problem. According to third-party model listings, the system carries a 220,000-token context window, caps output at roughly 4,100 tokens and exposes an OpenAI-compatible API as a closed-weight model.
The idea is not entirely new: OpenAI added crop-and-zoom during reasoning to its o-series models in 2025, and Google has explored related agentic vision work. Elorian’s argument is that it built the entire model around the capability rather than bolting it on. CEO and co-founder Andrew Dai traced the problem to a board-game night, when a frontier model could not say who was winning from a photograph.
“Each have their own visual language, each have their own iconography, symbols. To an adult, it’s quite easy to understand what’s going on with minimal text, but these models aren’t able to intuit what the symbols mean.”
He added: “Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can’t AI models?”
The benchmark ledger: four wins, one tie, six losses
Against GPT-6 Astra, Glimpse takes robotic manipulation events on WGO Bench (42.7 versus 39.4), open-vocabulary counting on FSC-147 (83.36 versus 82.02), short-form video captioning on UGC-VideoCap (82.66 versus 80.32) and scientific chart reasoning on CharXiv-R (92.9 versus 92.1) — although Claude Opus 5.5 leads that last test at 93.4. It ties on embodied spatial reasoning in ERQA at 83.75.
The losses are equally concrete: fine-grained visual reasoning on BabyVision (95.36 versus 97.68), grouped video comprehension on Video-MME v2 (56.68 versus 64.06), product-image counting on CountQA (82.66 versus 86.42), plus DREAM-1K, TOMATO and HSS. That final result is the most telling. HSS is a social-comprehension benchmark Elorian itself built with Scale AI and presented on 8 October; there, GPT-6 Astra tops the field at 53.6%, humans reach 93.1%, and Glimpse manages 46.4%.
Team, funding and the road from seeing to acting
Elorian emerged from stealth with a $55 million seed round at a $300 million valuation, co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. Dai previously worked at Google Brain and DeepMind, co-led pre-training for PaLM 2 and headed data for Gemini. Co-founder Yinfei Yang arrived from Apple and Google Research, while third co-founder Seth Neel was a Harvard professor.
Initial customers are organisations sitting on large video archives: robotics firms, scientific laboratories, media and creative companies. Elorian cites tagging robot teleoperation footage as an early application, a domain that matters as humanoid machines continue to rely on remote human operators. The company concedes that calibrated reconstruction and direct robot control remain open problems.
There is also a discrepancy over access. Elorian’s blog states that Glimpse is available today through its API and chat interface, while reporting on the launch indicates the chat interface requires a waitlist.
“The world’s best visual reasoning model.”
That is Dai’s stated target for the coming year, with a longer-term ambition of moving from seeing to doing.