What Glimpse is
Elorian released Glimpse on 16 October as a research preview. The company describes it as a visual thinking model: one that reasons about the physical world from what it sees, tracking objects through motion, reasoning from viewpoints other than the camera’s, and predicting physical consequences.
The design claim in the announcement is that the model brings visual representations into the reasoning process and can go back to them mid-thought, rather than converting an image to tokens once at the start and reasoning only over text thereafter. It also uses multimodal tools and can coordinate subagents.
Elorian says Glimpse was post-trained on an open-weights base model by a small team with a few hundred GPUs. It does not disclose a parameter count, saying only that the model is substantially smaller and cheaper than those it is compared against.
The price
Uncached, Glimpse costs $0.60 per million input tokens and $3.00 per million output tokens through Elorian’s API, with a chat interface on the same platform.
The company sets those rates against the field directly. They are 94% below GPT-6 Astra’s $10 and $50, 85% below Claude Opus 5.5’s $4 and $20, 70% below GPT-6.1 Sol’s $2 and $10, and 20% below Gemini 3.8 Flash’s introductory rates. Those comparisons are Elorian’s, drawn from published list prices.
The benchmarks, including the losses
The table in the announcement is unusually candid about where the model loses. Against GPT-6 Astra, Glimpse is ahead on four evaluations: WGO Bench, on manipulation events, at 42.70 against 39.42; CharXiv-R, on chart reasoning, at 92.90 against 92.10; UGC-VideoCap, on short-form video captioning, at 82.66 against 80.32; and FSC-147, on open-vocabulary counting, at 83.36 against 82.02. On ERQA, an embodied reasoning test, the two are level at 83.75.
On the rest — BabyVision, HSS, DREAM-1K, TOMATO, Video-MME v2 and CountQA — Glimpse trails. It does not lead every category even where it wins one: Opus 5.5 scores higher than both on CharXiv-R at 93.40, and Sol is ahead on Video-MME v2 with subtitles at 65.27.
All of these are Elorian’s own numbers, run by Elorian, on a model Elorian built. The usual caveat applies with more force than usual here, because the claim being made is about a price-performance ratio rather than a capability ceiling, and ratios are easier to construct favourably than a win on a saturated benchmark.
What to watch
Whether an independent index picks the model up. Elorian’s argument is that a small team with a few hundred GPUs, starting from someone else’s open weights, can reach the frontier on a narrow slice of visual reasoning at a fraction of the price. That is a claim about what post-training buys, and it only becomes interesting once someone other than Elorian measures it.
