Glimpse revisits images while reasoning, but Elorian measured its lead

Startup Elorian has introduced its first model, Glimpse, as a research preview. Published information says API access is available while the chat interface requires a waitlist. The company built the model around the idea that a system should repeatedly inspect the original visual source rather than reason only from a one-time conversion into tokens.

Glimpse reopens images and video while forming an answer

A mechanism called visual scratch space lets the model crop an image region, select a specific video frame or run code for measurement and comparison during reasoning. Elorian is targeting tasks where details, spatial relationships or the progression of a scene cannot be captured reliably in one text description.

The company says a small team trained the model on a few hundred GPUs and that it is substantially smaller than the frontier models used for comparison. Pricing is $0.60 per million input tokens and $3 per million output tokens.

Companies with large video archives stand to benefit first

Elorian names robotics companies, scientific laboratories, media groups and creative businesses as initial targets. For them, returning to specific frames may be more useful than producing a generally better video caption. It can tie a conclusion to a particular part of the recording and focus computation on the relevant moment.

The technique itself is not entirely new. TechCrunch notes that OpenAI added crop and zoom during reasoning to its o-series models in 2025. Elorian's bet is that this process should sit at the model's foundation rather than appear as an auxiliary tool.

The team selling the model still controls the scorecard

In Elorian's own evaluations, Glimpse beats GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual tasks. Those public claims are not yet independent verification, and the research preview limits how many outsiders can reproduce them.

It is also unclear how token pricing translates to video workloads, where the number of processed frames and repeated returns to the source can alter both total cost and latency. A price list without a representative workload tells only part of the story.

Independent tests on long video will determine the lead

The strongest signal will be reproducible testing by teams outside Elorian. It should measure not only accuracy on charts and short clips but also cost, latency, frames used and failure modes on longer or ambiguous scenes.

The company eventually wants to move from seeing to acting, including controlling robots. Before that step, Glimpse must show that visual scratch space works beyond selected examples and that revisiting the source produces more reliable decisions rather than merely a longer reasoning trace.

Lilith's verdict

Glimpse can return to a suspicious video frame and zoom in. It now needs an outside referee to inspect the scoreboard before the model is handed a robot.