Elorian launches Glimpse model with visual scratch space reasoning

Elorian released Glimpse, a vision model that re-queries the source image during reasoning — cropping, pulling video frames and running measurement code in a "visual scratch space" — instead of tokenizing an image once and reasoning from memory. Elorian claims Glimpse beats GPT-6 Astra on chart reasoning and video captioning despite being a much smaller model trained on a few hundred GPUs, and prices API access at $0.60 per million input tokens and $3.00 per million output tokens. The company positions the model as a cheaper, more accurate alternative to frontier multimodal models for dense visual workloads such as charts, diagrams, video and robotics feeds.

TechCrunch https://techcrunch.com/2026/10/16/elorians-first-ai-model-wants-to-reason-with-images-not-just-about-them/ Elorian launches Glimpse model with visual scratch space reasoning Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Glimpse keeps re-querying the source image mid-reasoning—cropping, pulling video frames, running measurement code in a "visual scratch space"—rather than tokenizing an image once and reasoning from memory, and it claims to beat GPT-6 Astra on chart reasoning and video captioning while being a much smaller model trained on a few hundred GPUs. At $0.60/$3.00 per million tokens in/out, it undercuts mid-tier models, so if you're processing dense visual data charts, diagrams, video, robotics feeds it's worth benchmarking as a cheaper, more accurate alternative to running frontier multimodal models on the same workloads. Glimpse exposes API access to a vision model that can re-inspect visual inputs during reasoning by cropping, pulling video frames, and running measurement code, priced at $0.60/M input tokens and $3/M output tokens. For teams building on image or video data, the practical shift is that some tasks currently handled with custom CV pipelines or human review—chart interpretation, short video captioning, board-state understanding—may become feasible through a general model, but at a cost tier above lightweight multimodal models.