Elorian released Glimpse, a model that re-reads images mid-answer

A visual model that writes one caption and stops will miss the number on the second-to-last bar of the chart. Elorian, a Palo Alto startup, released Glimpse, a model that it says reads an image, zooms into a region, extracts a video frame, or runs code to measure the picture, then rereads before it finishes the answer. The company calls this a visual scratch space, and the mechanism is the whole announcement.

Elorian is new and small. The company says its team post-trained Glimpse from an open-weights base model on a few hundred GPUs. Its founder Andrew Dai was a researcher at Google DeepMind, and TechCrunch reports Elorian has raised $55 million at a $300 million valuation from backers including Menlo Ventures, Altimeter Capital, Nvidia, and Jeff Dean. The model is in research preview with API access and a waitlist for the chat interface.

What the vendor measured

Elorian's own benchmarks give seven numbers. Glimpse scores 92.90 on CharXiv-R, a chart-reasoning test, and 94.33 on BabyVision, a fine-grained vision test, which the company says beats GPT-6 Astra on both. On UGC-VideoCap for short-form video captioning, Elorian reports 82.66. On WGO-Bench, which tests manipulation planning, it reports 42.70, and on ERQA, which tests embodied reasoning, 83.75. Open-vocabulary counting on FSC-147 is 83.36, and event understanding on E.T. Bench is 63.55. These are the vendor's own figures, run by the vendor, so an independent read would be the first thing a buyer should wait for.

The price gap

Glimpse is priced at $0.60 per million input tokens and $3.00 per million output tokens. Elorian states this is 94 percent cheaper than GPT-6 Astra on both. TechCrunch reports the pricing sits below Anthropic's Sonnet and Sol tier and above Haiku and Luna. A sustained price gap of this size usually means the model is smaller than the frontier it compares to, or trained on a tighter data budget, or both. Elorian does not state the parameter count.

What Elorian does not claim

The announcement lists limits in its own voice. Direct action generation is future work. Calibrated reconstruction, which would let a robot act on the picture, is work ahead. Robot control is the intended end state and the current release is not it. The release pitches four initial uses: robotics, scientific labs, media, and creative companies. On a research preview, chat is waitlist only.

The pattern this fits is a model that treats an image as a workspace a reasoner can edit, rather than a sentence to describe. If the benchmark wins hold up in independent testing, a team building a chart reader, a measurement tool, or a video review assistant has a new price point to compare. If they do not, the lesson is that vendor benchmarks are vendor benchmarks and the only test that matters is your own.

Source

Elorian, Introducing Glimpse. TechCrunch, Aditya Mehta, Elorian wants its model to reason with images.