Elorian, a Palo Alto AI startup, has released its first model. Glimpse is a visual reasoning model that keeps working with images and video while it reasons, rather than turning them into text once and reasoning from that, and it is available now through Elorian's API and chat interface. It costs $0.60 per million uncached input tokens and $3.00 per million output tokens. Elorian says that on multimodal benchmarks Glimpse performs at the level of GPT-6 Astra and Claude Opus 5.5, sometimes exceeding them, while being a substantially smaller and cheaper model. Elorian's launch post is dated October 16, and TechCrunch, which reported the launch as an exclusive, published its story at 00:01 Beijing time on October 17.

How Glimpse works

Elorian calls Glimpse "a new category of visual thinking model," one that reasons about the physical world from what it sees: following objects as they move and as the viewpoint changes, working out how parts fit together, and connecting an observation to what happens next. The company says the model can revisit and refine its internal visual representations as it thinks. TechCrunch describes what that looks like in practice: mid-reasoning, Glimpse can crop into a region, pull specific frames from a video or run code to measure what it is looking at, a mechanism Elorian calls a "visual scratch space." Elorian's chat platform shows that scratch space while the model works.

Glimpse is also trained end to end as an agent that uses multimodal tools to write code, transform images and watch videos, and Elorian says it can spawn subagents that pass text, visual artifacts and internal representations to each other while working on the same problem. The company is pitching it at robotics, scientific work and the creative industries. Its examples include labeling robot teleoperation footage, planning the steps for a vision-language-action robot policy, captioning short videos in detail and counting objects through a moving camera shot.

Elorian's benchmark numbers

These scores are Elorian's own, from the table in its launch post. We haven't run these tests. Elorian says it ran every model at its highest thinking effort with its native agent harness, adding Code Interpreter for the GPT, Claude and Gemini models, with some exceptions on the video evaluations. By its numbers, Glimpse beats GPT-6 Astra on four evaluations and ties it on a fifth:

  • WGO Bench, robot manipulation events: 42.70 vs 39.42
  • CharXiv-R, scientific chart reasoning: 92.90 vs 92.10
  • UGC-VideoCap, short-form video captioning: 82.66 vs 80.32
  • FSC-147, open-vocabulary counting: 83.36 vs 82.02
  • ERQA, embodied spatial reasoning: 83.75 for both

The same table shows Glimpse behind Astra on the other six evaluations where Astra has a score, including BabyVision (95.36 vs 97.68), TOMATO (93.60 vs 96.02) and Video-MME v2 with subtitles (56.68 vs 64.06). Against Claude Opus 5.5, Glimpse scores higher on seven of the nine evaluations where both models have results, and lower on CharXiv-R (92.90 vs 93.40) and HSS scene understanding (46.36 vs 49.81).

Price and availability

Elorian says its $0.60 and $3.00 rates are 94% lower than GPT-6 Astra's, 85% lower than Claude Opus 5.5's, 70% lower than GPT-6.1 Sol's and 20% lower than Gemini 3.8 Flash's introductory rates. TechCrunch places it between tiers: cheaper than mid-level models such as Sonnet and Sol, but more expensive than lightweight models such as Haiku and Luna. Elorian also notes that the close inspection behind Glimpse's answers takes time and tokens, and says token efficiency is its next focus.

Glimpse is also listed on OpenRouter. The OpenRouter page shows a cache-read price of $0.06 per million tokens, a context window of 262,144 tokens and up to 256,000 output tokens, with text, image and video input and text output. Together is the only provider listed there for now.

Elorian says it built Glimpse as a small team with only a few hundred GPUs, post-trained on "an open weights base model." It doesn't say which base model it used. Elorian hasn't released Glimpse's own weights, so for now access is through the API, the chat interface and OpenRouter.

The company behind it

According to TechCrunch, Elorian raised a $55 million seed round at a $300 million valuation before it had launched a product. The company came out of stealth six months ago with a round co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. Its CEO, Andrew Dai, is a former Google DeepMind researcher. TechCrunch reports that Glimpse is launching as a research preview and that Elorian's first targets are companies with large amounts of video, such as robotics firms, scientific labs, and media and creative companies. Dai told TechCrunch the company aims to build "the world's best visual reasoning model" next year. In its own post, Elorian says its next steps include multimodal-native pretraining, models that can both understand and generate, and direct action outputs for controlling robots.