Elorian Glimpse

What visual thinking means here

Elorian describes a model that revisits visual evidence during reasoning, with an agent trained to inspect images, use code and coordinate work. The name expresses the company’s approach; it does not establish that other multimodal models lack visual reasoning or that every hosted endpoint includes its agent tools.

Watch the train example and check the evidence

In the announcement’s recorded trace, Glimpse follows one window edge and compares its displacement over time to infer that the train is slowing. Watch the source clip here, then read the timestamped trace at Elorian. The clip and the written explanation serve different roles.

Choose an access surface

Elorian links a chat platform and API; the documentation entry required sign-in when checked on October 17. This does not establish access for every account. Together lists dedicated deployment, with endpoint details supplied with the deployment. OpenRouter and Vercel publish the exact route ID elorian/glimpse.

• Glimpse API specifications

  Compare hosted limits, prices and a documented request shape.

• Try the Elorian platform

  Check current chat access and account requirements on the publisher’s platform.

• AI Gateway lists 220,000 context tokens and a 4,096-token output ceiling.
• OpenRouter’s current Together endpoint lists 262,144 context tokens and 256,000 maximum completion tokens. These are that offer’s advertised limits, not a measured usable budget.
• Use the selected host’s current multimodal schema. Do not assume that its response includes Elorian’s visual scratch space or the launch evaluation’s max-effort agent.

Read token prices as one part of task cost

Elorian quotes $0.60 per million input tokens and $3 per million output tokens. Vercel and OpenRouter list the same base rates plus $0.06 per million cached-input tokens. Cache terms belong to the chosen host. Dense visual inspection, retries and a longer answer can change the final bill; a lower unit rate does not establish a cheaper correct result.

What the published evaluations establish

Elorian’s table puts Glimpse ahead of Astra on four of eleven shared evaluations, tied on one and behind on six. The table uses different agent harnesses and some different thinking settings. Read it as a vendor comparison under stated conditions, rather than a universal ranking. The demonstrations do not establish reliable game coaching, viral-video creation or direct robot control.

• GPT-6 Astra

  Compare the broader model’s scope and documented workflow boundaries.

Team background and the earlier financing

Elorian identifies Andrew Dai as CEO and a former Google Brain/DeepMind researcher, and Yinfei Yang as its multimodal architect who led Apple MM1 and MM1.5. Wilson Sonsini says the company announced a $55 million seed round on April 9, 2026. That financing predates Glimpse’s October launch and is not evidence of production adoption or model quality.

Try a bounded visual task before adopting it

Start with clips or images you are allowed to upload and a reference answer. Ask for the decisive region, object identity or timestamp, then check it yourself. For a model comparison, keep inputs, tools and success criteria comparable and record retries and total billed use.

• Agent evaluation tools

  Find ways to track datasets, traces and human review.

• For counting, include moving-camera footage, occlusion and lookalike objects.
• For captions, check event order and separate visible actions from inferred intent.
• For physical reasoning, review a proposed consequence before using it to guide a real action.

Useful first jobs

• Evaluate image and chart interpretation
• Annotate short videos and inspect event order
• Count and track objects with a reviewed reference answer