Elorian's Glimpse model is cheaper than GPT-6 Astra and outperforms it in some tests

Summary

Elorian, a Palo Alto-based startup founded by former Google researchers, has introduced Glimpse. This visual reasoning model repeatedly examines images, videos, and code outputs during inference. The company claims that Glimpse outperforms GPT-6 Astra in certain benchmarks and offers token prices that are approximately 94% lower.

Why it matters

This demonstrates that a small team with just a few hundred GPUs can compete with or even surpass leading labs in narrow visual tasks, creating genuine price pressure on incumbents. The model's 'visual thinking' loop (cropping, zooming, frame extraction, and executing code during inference) targets practical workloads in robotics, scientific imaging, and media. The $55 million seed round at a $300 million valuation, backed by Nvidia and Jeff Dean, reflects investor interest in compact, task-focused multimodal models.

Highlights

• Glimpse outperforms GPT-6 Astra in WGO Bench (42.7 vs. 39.4), UGC-VideoCap (82.7 vs. 80.3), and CharXiv-R (92.9 vs. 92.1). However, it lags behind in Video-MME v2 (56.7 vs. 64.1) and social understanding HSS (46.4%) tests.
• Pricing is set at $0.60/$3.00 per million tokens. This represents a 94% discount compared to Astra's $10/$50 pricing, targeting robotics, laboratory, and media customers.
• The model was trained on an undisclosed open-weight base using 'only a few hundred GPUs.' The parameter count has not been disclosed.

Critical take

All benchmarks are self-reported by Elorian. Additionally, the model's iterative examination of images may increase total token consumption, potentially negating the per-token price advantage in practice. Notably, the company obscures this figure from its own comparison table.