Elorian’s Glimpse Lists API Rates 94% Below GPT-6 Astra, With Mixed Results

Elorian released Glimpse on Friday, Oct. 16, as a research preview with listed API rates 94% below GPT-6 Astra’s, while its benchmark results split among four wins, one tie and five losses.

The model is available through an API and can return to visual evidence as a task develops. It can focus on selected image areas, inspect video at denser frame rates, and run code to measure or compare visible details. Those tools let it revisit parts of an image or video while working through a request.

The API lists Glimpse at $0.60 per 1 million input tokens and $3 per 1 million output tokens, compared with Astra’s $10 and $50 rates. The listed price per token is 94% lower for both input and output. Those rates do not show what a completed task will cost: close visual inspection takes time and tokens, and token efficiency is an area for further work.

Glimpse scored higher on four of the 10 comparisons. It posted 42.70 versus 39.42 on WGO Bench for end-to-end manipulation events, 92.90 versus 92.10 on CharXiv-R for chart reasoning, 82.66 versus 80.32 on UGC-VideoCap for short-form video captioning, and 83.36 versus 82.02 on FSC-147 for open-vocabulary counting.

Both models scored 83.75 on ERQA for embodied spatial reasoning. Astra scored higher on five tests: HSS, 53.6 to Glimpse’s 46.36; DREAM-1K, 79.05 to 76.04; TOMATO, 96.02 to 93.60; Video-MME v2, 64.06 to 56.68; and CountQA, 86.42 to 82.66.

The comparison used each model’s native agent harness and highest thinking-effort setting, with exceptions for some video evaluations. The HSS results also include scores from research papers.

Potential uses include robotics, scientific research, media and creative work. The expanded scorecard follows earlier coverage of Glimpse’s first two visual evaluation wins.