LLMs are either not capable or not affordable for visual reasoning tasks at scale. The Elorian team - one of Jeff Dean’s best reports who led Gemini, an Apple LLM lead, xAI lead, and a Harvard professor - just launched an Astra-level visual reasoning model, Glimpse, for 17x cheaper.... Watch it break down the flying Tesla video physics here. Extremely critical for video inputs in manufacturing, robotics, gaming, content moderation and more.show more

Available on OpenRouter:

Fortunate to have led the seed round to partner with the company alongside @strikervp and @AltimeterCap.

this is all still insanely slow tbh

run it on a video where the angle changes mid-shot and something walks out of frame and returns, if it keeps track of that then the claim actually holds

Try gemini, it works very good for me in explaining physics from video

The model spotted the physics. I spotted 'I led the seed round.'

i hate when these models dont work for me. img in ref :

Point it at a whiteboard someone photographed sideways from the back row and I'm sold...

everyone's fixating on the 17x cheaper number. the scarier part for the labs: a tiny team on a few hundred GPUs just beat the frontier on charts and video. visual reasoning might be the first frontier capability that doesn't need frontier money.

Lol when the elorian looks like it was fully vibed with Claude

every launch is astra-level until you throw a moving object at it lol

no opensource no value

does the 17x hold on real traffic?

Looks cheap on cost at least Need to test for rest

17x cheaper is wild if it holds up

Been using ffmpeg for deep video analysis with frontier models, it works flawless >90% of time... But still have to give it proper instructions on what to do and what not to.

先看杂乱场景下能不能稳定识别，再谈大规模用。

the 17x cost reduction only matters if accuracy stayed competitive with astra. what did the benchmark gaps look like?

Lower costs could change everything.

17x cheaper is the part that matters. Video understanding only goes mainstream once the price drops like that.

gemini, apple and xai leads on one visual reasoning bet says where they think the gap is

if it can follow what's happening across a whole video and not just one frame, that's the first ai i'd trust with my edits lol

Cost is the unlock everyone underestimates in AI. A visual reasoning model that is 17x cheaper turns video from a demo feature into default infrastructure for factories, robotics, etc. Pricing > Benchmarks!

Using the flying Tesla Roadster stunt video as the universal benchmark to test if a brand new AI model understands real-world physics is peak 2026 tech culture.

@deedydas what did the 17x refer to in the cost comparison?

Testing it on a flying Tesla video is a choice. Respect.

Dense video reasoning at 17x lower inference cost is the exact unlock robotics and manufacturing need. Processing temporal frames through frontier multimodal models has always been bottlenecked by token economics, not architecture.

visual reasoning that's actually affordable at scale is the part everyone skips over

Can you compare this to the costs and speed of 3.8 Flash with agentic video understanding?

getting visual reasoning 17x cheaper on Glimpse means nothing if it still chokes the moment you feed it thirty frames a second.

17x cheaper per what though, per frame or per minute of video? for robotics the latency number would matter more to me

I'd feed it my own line footage, not the demo clip, before pricing in the 17x.