Elorian Launches Glimpse, a Visual Model Priced at $0.60 per Million Tokens, Scored Only by Elorian

Elorian, a Palo Alto startup that raised a $55 million seed round at a $300 million valuation before it had a product, launched its first model, Glimpse, as a research preview on October 16. Glimpse reasons over images and video by going back to the source pixels mid-thought, and Elorian says it beats GPT-6 Astra on scientific chart reasoning and short-form video captioning. Every figure comes from Elorian's own evaluations.

What Glimpse does differently

Most multimodal models turn an image into tokens once and reason over those. TechCrunch reported on October 16 that Glimpse can return to the source during reasoning: crop regions, pull frames from a video, and run code to measure or compare. Elorian calls this a "visual scratch space."

Andrew Dai, Elorian's chief executive and a former Google DeepMind researcher, told TechCrunch the goal is "the world's best visual reasoning model" next year. His example of the gap: "Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can't AI models?"

Together AI's model page adds that Glimpse takes text, image and video in and returns text. It describes post-training on an open-weights base model that it does not name. It also lists subagent coordination, where subagents pass text, visual artifacts and internal representations to one another.

Elorian's benchmark numbers

Together AI's page lists scores that it labels as Elorian-reported, at maximum thinking effort. No independent lab has run them, and the page names no scores for GPT-6 Astra to compare against.

Benchmark	Area	Elorian-reported score
CharXiv-R	Scientific charts	92.90
TOMATO	Temporal reasoning	93.60
BabyVision	Visual perception	95.36
ERQA	Robotics	83.75
UGC-VideoCap	Video captioning	82.66
Video-MME v2 (with subtitles)	Video	56.68

The page lists more, including 42.70 on WGO Bench end-to-end manipulation and 76.04 on DREAM-1K video captioning. The claim that matters commercially is the GPT-6 Astra comparison, which TechCrunch says rests on Elorian's evaluations and was not independently verified.

Price, access and who is paying

TechCrunch gave a price of $0.60 per million input tokens and $3 per million output tokens. That sits above lightweight models such as Haiku and Luna and below mid-level ones such as Sonnet and Sol, according to the same report.

Together AI lists no per-million-token price. It says Glimpse is offered through dedicated deployment, with endpoint details supplied on deployment, and carries the status "Launching soon." The page gives a release date of October 15, a day before TechCrunch's launch date, and the discrepancy is unexplained.

The round was co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, TechCrunch said. Nvidia and Jeff Dean, Google's chief scientist, also took part. Elorian's own site lists Striker Ventures, Menlo Ventures and Altimeter, with participation from NVIDIA and 49 Palms.

Customers and what is missing

Elorian's first targets are robotics firms, scientific labs, and media and creative companies with large video libraries. Longer term the company wants to move from seeing to acting, including controlling robots directly, TechCrunch reported.

Elorian says a small team trained Glimpse on a few hundred GPUs and that it is smaller than the frontier systems it is compared with. No parameter count is published. Neither TechCrunch nor Together AI states a context window, and there is no technical report or licence for the weights, if any are released.

Together AI's own deployment is the next milestone to watch. Its page promises an email to people who sign up when the endpoint goes live, and gives no date.

For comparison with other model launches on this desk, see the Gemini 4 Argon pricing breakdown and the HybridAgent paper's unclear baseline.