Elorian Glimpse: a $0.60/$3 visual-reasoning API that beats GPT-6 Astra on video captioning — and where it still trails
Elorian's first model, Glimpse, is a research-preview API and chat UI that re-inspects images and video mid-reasoning. It is priced below Gemini 3.8 Flash, leads on chart, counting and short-video captioning evals, but trails Astra on long video QA and temporal reasoning. Here is when it is worth testing.
Elorian, the Palo Alto lab that raised a $55 million seed round before shipping anything, released its first model on October 16, 2026. Glimpse is a “visual thinking” model: instead of turning an image into tokens once and reasoning in text, it can crop into a region, re-watch part of a video at a denser frame rate, or run code to measure what it sees while it is still thinking. Elorian calls that working area a visual scratch space.
The builder question is simple: is a small lab’s research-preview model worth adding to a vision or video pipeline? For some workloads, the published numbers say yes.
Price and access
| Input (per 1M, uncached) | Output (per 1M) | |
|---|---|---|
| Glimpse | $0.60 | $3.00 |
| Gemini 3.8 Flash (introductory) | $0.75 | $3.75 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| GPT-6 Astra | $10.00 | $50.00 |
Prices are from Elorian’s own comparison table. Glimpse is available through an API and a chat interface on the Elorian platform, and the support page points enterprise buyers to sales@elorian.ai. The API documentation sits behind sign-in, so we could not confirm model IDs, context length or rate limits. Check those in the console before you plan production traffic. If you are comparing it with Google’s option, see our note on Gemini 3.8 Flash GA pricing.
Where it leads, and where it does not
All scores below are Elorian’s own runs at the highest thinking effort, not independent tests.
Ahead of GPT-6 Astra: WGO Bench manipulation events (42.70 vs 39.42), CharXiv-R chart reasoning (92.90 vs 92.10, though Opus 5.5 scores 93.40), UGC-VideoCap short-form captioning (82.66 vs 80.32, the best in the table), and FSC-147 open-vocabulary counting (83.36 vs 82.02, also the best). It ties Astra on ERQA embodied spatial reasoning at 83.75.
Behind: Video-MME v2 (56.68 vs 64.06 for Astra and 65.27 for Sol), HSS scene understanding (46.36 vs 53.6), DREAM-1K dense captioning (76.04 vs 79.05), TOMATO temporal reasoning (93.60 vs 96.02), and CountQA product counting (82.66 vs 86.42).
The pattern is clear. Short clips, charts and counting objects in a scene favor Glimpse. Long-video question answering and fine temporal reasoning still favor the frontier models.
The cost catch: tokens per task
Elorian says plainly that looking closely “takes time, and tokens,” and that token efficiency is its next focus. A low per-token price can be eaten up if the model zooms and re-watches a lot. Its one per-task figure is a robot-planning episode on a LIBERO task with the π0.5 policy: $0.035 with Glimpse versus $0.249 with Astra, and both completed the task. Measure cost per completed task on your own clips, not cost per million tokens.
Who should test it now
- Teams captioning or tagging large short-form video libraries, or labeling robot teleoperation data. Elorian names robotics, science labs and media as its first targets.
- Pipelines that count objects or read scientific charts, where it matches or beats far pricier models.
- Anyone already testing video-understanding APIs such as Gemini Robotics ER 2, as a cheaper comparison point.
Treat it as a research preview. Elorian is a small lab with no ticket queue (support is email-only), and it says it post-trained Glimpse on an open-weights base model with a few hundred GPUs. Run it next to your current model on your own data before you switch. For document-style extraction you can host yourself, our Qwen3-VL extraction pipeline guide is still the self-hosted baseline.
Source: Elorian — Introducing Glimpse