Elorian Glimpse: Visual Reasoning, API Access and Task Costs

Sign in or register on Elorian’s platform and check the access enabled for your account. The announcement offers API and interactive chat routes. For an API integration, use the endpoint, model name and authentication instructions shown in your account.

For a first chat task, upload a short clip or image you can independently check. Ask one question, inspect the answer and its visual evidence, then change one ambiguity at a time. Preserve the source file and question so another model can receive the same task.

Elorian presents visual inspection and revisiting regions or video intervals as part of reasoning. Its launch examples cover identity tracking, viewpoint changes, counting and physical consequences. Some are recorded model traces; others are source media with reviewed reference answers. Keep that distinction when judging a demonstration.

The launch benchmark table is Elorian’s evaluation with specified thinking effort and harnesses, not an independent deployment test. A useful comparison for your work is whether each answer identifies the relevant chart region, event interval or object and survives checking against the original.

Robot planning examples use Glimpse around a separate robot policy. They do not make it a general video-generation endpoint or a released direct-action controller.

The announced research-preview price is $0.60 per million uncached input tokens and $3.00 per million output tokens. For assumed usage of 100,000 input and 10,000 output tokens, the model-token bill is $0.06 + $0.03 = $0.09. This is an arithmetic example, not a measured image or video request.

Visual inspection and further reasoning can increase usage. Record the service’s returned token usage for all calls and retries; a video’s playback duration does not by itself determine billable tokens. Use any separately listed tool or service charges from your account’s billing terms.

Create a small review set with known answers. For a chart, retain the values and labels; for a video, mark the event interval; for counting, define the qualifying objects before looking at the model result. Track accepted answers, missed evidence, total latency and the complete bill.

Run the same source media and question through your existing model. Keep the thinking setting and tool access with each result. An answer that sounds plausible but points at the wrong event is a failed task even if its token bill is small.

Elorian says Glimpse was post-trained on an open-weight base. The announcement does not provide a Glimpse checkpoint download, exact parameter count or self-hosting recipe. Use the hosted access and published rates for this model rather than assigning it a GPU tier from the base-model description.