Elorian launches Glimpse, a visual-thinking model that revisits images and video frames mid-reasoning; $0.60/$3 per M tokens, ahead of GPT-6 Astra on four vendor-run evals

Background

Most multimodal models look at an image once, tokenize it, then reason in text, which breaks on motion, viewpoint changes and precise counting. On Humanity's Sixth Sense (built by Elorian with Scale AI), humans score 93.1% while GPT-6 Astra scores 53.6%. Elorian, founded by ex-DeepMind researcher Andrew Dai, shipped its first model Glimpse on Oct 16, 2026.

How it works

Glimpse reasons in a "visual scratch space": it crops and zooms regions, rewatches short intervals at denser frame rates (8 fps to ~14 fps in one published trace), and runs code to measure. It is trained end-to-end to use multimodal tools and can spawn subagents that share text, visual artifacts and internal representations. Elorian says a small team post-trained it on an open-weights base using a few hundred GPUs.

Benchmarks (vendor-reported)

Ahead of GPT-6 Astra on WGO Bench (42.70 vs 39.42), UGC-VideoCap (82.66 vs 80.32), FSC-147 (83.36 vs 82.02) and CharXiv-R (92.90 vs 92.10; Opus 5.5 leads at 93.40); tied on ERQA (83.75). Behind on HSS (46.36 vs 53.6), Video-MME v2 (56.68 vs 64.06), TOMATO and BabyVision. No independent evals yet.

Getting started

API on the Elorian platform at $0.60/$3 per M input/output tokens. Good fits: labeling and segmenting teleoperation data, dense video captioning, counting in site or factory footage, and planning for VLA robot policies ($0.035 per LIBERO episode vs $0.249 with Astra). Benchmark it on your own clips before switching.