Glimpse
Elorian's first AI model, which adopts a vision-native reasoning architecture in which intermediate representations during reasoning remain in visual form rather than being translated into text tokens.
Timeline (last 90 days)
Glimpse is Elorian's first model, released as a research preview with both an API and a chat interface.
Glimpse may not be the model that ends up being adopted, but its API is worth trying while leaving room for verification.
Chart reasoning benchmarks are themselves easy to cherry-pick, and a test set built around "a model that crops its own images" will naturally flatter the model built to crop its own images unless the evaluation is carefully blinded.
A model that re-crops images and runs code mid-reasoning does more work per answer and may have higher latency.
Glimpse is a research preview, which in practice means the API may have breaking changes, rate limits, and rough edges, and no downloadable weights are mentioned, so plan on an API-only availability.
Glimpse's announcement does not explain how images and video frames are tokenized or how billing works, and this detail is critical to the actual cost of a vision model.
Glimpse's claim of "surpassing GPT-6 Astra" covers only two tasks—scientific chart reasoning and short video description—and comes from the lab's own announcement, with no third-party benchmark yet.
According to TechCrunch, a small team trained Glimpse with a few hundred GPUs, while frontier systems cost orders of magnitude more.
Glimpse's visual draft space architecture requires the model to have two capabilities: metacognition to judge whether the current information is sufficient, and an interface to translate operations such as crop coordinates and frame timestamps into executable instructions; the cost is a significant increase in inference latency and computation.
The 12-row benchmark table released by Elorian shows that Glimpse won 4 rows, tied 1, and lost 6 against GPT-6 Astra, with Astra having no score in 1 additional row.
24 more timeline events
All Facts (20)
Glimpse's visual draft space architecture requires the model to have two capabilities: metacognition to judge whether the current information is sufficient, and an interface to translate operations such as crop coordinates and frame timestamps into executable instructions; the cost is a significant increase in inference latency and computation.
50%UnverifiedGlimpse may not be the model that ends up being adopted, but its API is worth trying while leaving room for verification.
50%UnverifiedChart reasoning benchmarks are themselves easy to cherry-pick, and a test set built around "a model that crops its own images" will naturally flatter the model built to crop its own images unless the evaluation is carefully blinded.
50%UnverifiedA model that re-crops images and runs code mid-reasoning does more work per answer and may have higher latency.
50%UnverifiedGlimpse is a research preview, which in practice means the API may have breaking changes, rate limits, and rough edges, and no downloadable weights are mentioned, so plan on an API-only availability.
50%UnverifiedGlimpse's announcement does not explain how images and video frames are tokenized or how billing works, and this detail is critical to the actual cost of a vision model.
50%UnverifiedGlimpse's claim of "surpassing GPT-6 Astra" covers only two tasks—scientific chart reasoning and short video description—and comes from the lab's own announcement, with no third-party benchmark yet.
50%UnverifiedAccording to TechCrunch, a small team trained Glimpse with a few hundred GPUs, while frontier systems cost orders of magnitude more.
50%UnverifiedThe 12-row benchmark table released by Elorian shows that Glimpse won 4 rows, tied 1, and lost 6 against GPT-6 Astra, with Astra having no score in 1 additional row.
50%UnverifiedGlimpse is priced on Elorian's own platform at $600,000 per million input tokens and $3,000,000 per million output tokens.
50%UnverifiedElorian claims that Glimpse's pricing is 94% cheaper than GPT-6 Astra's $10/$50.
50%UnverifiedOn the HSS (image + video) benchmark, Glimpse scored 46.36 and GPT-6 Astra scored 53.6, a gap of about 7 points between the two.
50%UnverifiedGlimpse is built on an unnamed open-weight base model and post-trained by a small team with only a few hundred GPUs.
50%UnverifiedGlimpse's model name on OpenRouter is elorian/glimpse; endpoint data shows the provider is Together, the context window is 262,144 tokens, it supports text, image, and video input and outputs text, and the cached input read price is $0.06 per million tokens.
50%UnverifiedDuring its thinking process, Glimpse repeatedly returns to the original image, and can crop regions, extract frames from video, and run code to measure, which Elorian calls a visual scratchpad.
50%UnverifiedTechCrunch positions Glimpse's price above lightweight models such as Haiku and Luna, and below mid-range models such as Sonnet and Sol.
50%UnverifiedOn a single LIBERO robot planning task, one episode costs $0.035 for Glimpse and $0.249 for GPT-6 Astra.
50%UnverifiedOn the CharXiv-R benchmark, Claude Opus 5.5 ranks first with 93.40, ahead of Glimpse and GPT-6 Astra.
50%UnverifiedThe officially given Glimpse use cases cover scientific chart analysis, video description generation, visual counting, and robot data annotation.
50%UnverifiedGlimpse's core capability is "returning to visual evidence," that is, zooming in on details, re-watching video segments, or tracking the same feature across frames during reasoning.
50%