Elorian Glimpse model rechecks images while it reasons

Elorian released its first AI model, Glimpse, on October 16, 2026, as a research preview with API access. Glimpse is built to look back at an image or video while it works out an answer, instead of turning the picture into text once and reasoning from that. It costs $0.60 per million input tokens and $3 per million output tokens, which Elorian says is 94% cheaper than OpenAI's GPT-6 Astra.

Elorian is a startup that raised $55 million before shipping any product, TechCrunch reports. Its backers include Striker Ventures, Menlo Ventures, Altimeter and Nvidia, according to the company's website. Andrew Dai is its co-founder and CEO.

How Glimpse looks at an image twice

Most vision models today convert an image into text-like data at the start, then reason in text. Elorian argues on its website that this is fragile and leads to made-up answers, or hallucinations. Glimpse instead can crop part of an image, pull out single video frames, and run code while it reasons. Elorian calls this a "visual scratch space," TechCrunch reports.

The company's launch post gives an example with a video of a bird. Glimpse first scans the clip at 8 frames per second. When it is unsure, it goes back and rescans a 0.9-second window at about 14 frames per second. Elorian sums up the approach in one line: Glimpse "questions its first impression, and returns to the visual evidence before it answers."

Dai explained the goal to TechCrunch with a furniture example. "Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can't AI models?"

Elorian says a small team trained Glimpse on "only a few hundred GPUs," starting from an open-weights base model. It has not said which base model, and it has not published the parameter count, context window or rate limits, OrcaRouter notes.

Benchmarks: four wins and six losses against Astra

Elorian's launch post compares Glimpse with GPT-6 Astra on 11 benchmarks. Glimpse wins four, ties one and loses six. All the numbers are Elorian's own and have not been checked by anyone else. RuntimeWire notes that the company's own table shows those losses.

Benchmark	Glimpse	GPT-6 Astra
WGO Bench	42.70	39.42
CharXiv-R	92.90	92.10
UGC-VideoCap	82.66	80.32
FSC-147	83.36	82.02
ERQA	83.75	83.75
BabyVision	95.36	97.68
CountQA	82.66	86.42
DREAM-1K	76.04	79.05
TOMATO	93.60	96.02
Video-MME v2	56.68	64.06
HSS	46.36	53.6

HSS, short for Humanity's Sixth Sense, is a test Elorian itself released nine days before Glimpse, OrcaRouter reports. OrcaRouter also reads the table as RuntimeWire does. Glimpse leads on perception tasks, where the model must look carefully, and trails on harder inference tasks such as HSS.

What it costs compared with other models

Elorian lists these prices per million tokens in its launch post. A token is a chunk of text or image data that a model reads or writes.

Model	Input	Output
Elorian Glimpse	$0.60	$3.00
Gemini 3.8 Flash (introductory)	$0.75	$3.75
GPT-6.1 Sol	$2.00	$10.00
Claude Opus 5.5	$4.00	$20.00
GPT-6 Astra	$10.00	$50.00

Astra's $10 input price matches what Amazon charged when GPT-6 Astra became generally available on Bedrock in September. Elorian also says one robot-planning task cost $0.035 with Glimpse and $0.249 with Astra.

What this means for developers

Glimpse is a cheap new option for vision-heavy work, but it is a preview from a young company. Test it on your own images and videos before you move anything to it.

• Pick tasks that match its strengths. Its clearest wins are scientific chart reading and short video captions, TechCrunch reports. Elorian also pitches robotics planning, data labeling and counting as early uses, for robotics companies, science labs and media firms.
• Check video work carefully. Glimpse trails Astra by more than 7 points on the Video-MME v2 video benchmark.
• Plan for missing details. With no published context window or rate limits, size a test workload small and measure latency yourself.
• Watch the HSS scores. Elorian built that benchmark, so an outside lab reproducing the results would mean more than any number in the launch post.

The chat app is waitlist-only for now. Developers can sign up for API access at platform.elorian.ai. Dai told TechCrunch that controlling robots directly is a longer-term goal, and Elorian lists it as future work.