Elorian launches its first model, Glimpse

A model that looks again

Elorian says Glimpse can inspect image regions, revisit video intervals and use code during reasoning. Its API pricing is $0.60 per million input tokens and $3 per million output tokens. Launch announcement

TechCrunch describes the release as a research preview. Original report

Strong results, uneven leadership

Elorian reports 92.90 on CharXiv-R, versus Astra’s 92.10 and Opus 5.5’s 93.40. Glimpse ran at maximum effort; Astra and Opus used maximum effort with Code Interpreter. These are vendor tests. Evaluation table

CharXiv helps explain what that comparison measures. Its authors assembled 2,323 scientific charts and distinguish reading individual chart elements from synthesizing several visual elements. Human experts curated the material. Success on that narrower test is useful evidence for document analysis; it does not establish general visual understanding. CharXiv repository

Seeing remains a separate challenge

Scale and Elorian’s October 7 Humanity’s Sixth Sense research tests the temporal, spatial, social and abstract information implied by scenes. Its human participants reached 93.1%, while the strongest tested model, GPT-6 Astra, reached 53.6% at maximum reasoning effort. Dynamic visual tools narrowed that gap without closing it. Those findings make scene interpretation a separate evaluation requirement for robotics or video workflows. HSS research

The underlying idea also has a history: OpenAI described native cropping, zooming and rotation during reasoning in April 2025. It warned of perception errors and redundant tool calls. Glimpse’s commercial test is therefore whether its implementation improves accuracy and total workflow cost on customers’ actual visual tasks. Thinking with images