Elorian Launches Visual Thinking Model Glimpse

Episode 1 · Ex-Gemini team's Elorian launches visual-thinking model Glimpse, claiming Astra-level performance (2026-10-17, 45 posts)

On October 17, Palo Alto startup Elorian released Glimpse, its first "visual thinking model," as a research preview with API available and chat waitlisted, priced at $0.6 per million input tokens and $3 per million output tokens. Founder Andrew Dai, a former DeepMind/Gemini researcher, claims the model matches GPT-6 Astra and Claude Opus 5 on key public vision-reasoning benchmarks (leading on 4 of them), was trained on only a few hundred GPUs at roughly 6% of Astra's cost, is 17x cheaper per token, and is smaller and cheaper than Gemini 3.8 Flash. Jeff Dean and Google research VP Milanfar both shared praise; testingcatalog reports a $55M seed round at a $300M valuation closed before launch. If the claims hold, Glimpse would approach frontier vision reasoning at a fraction of Big Tech's compute, warranting close third-party verification.

Confirmed

• Positioning: targets the weakness that models excel on single static images but fail when scenes move or viewpoints change; three core tasks—identity tracking of moving objects, cross-viewpoint reasoning, and predicting physical consequences. Target scenarios include factory floors, construction sites, warehouses, robot motion, and long video footage; the company also highlights complex counting and visual puzzles.
• Technique: a "visual scratch space" keeps visual representations inside the reasoning loop—the model can re-examine images, crop and zoom, pull specific video frames, and run code to measure and compare, "thinking while looking" like a human pinching to zoom a photo; official material also mentions multimodal tools and sub-agents, with end-to-end agentic training.
• Commercialization: the founder cites computer use, CAD, and multimodal data labeling as near-term applications; an analysis he reshared argues robotics is the biggest use case, e.g., robots doing a visual check every 2 seconds.
• Team & funding: core members come from Google Gemini and Apple MM-1 (member yinfeiy says a decade of foundation-model work at Google and Apple); backers disclosed by Andrew Dai include Menlo Ventures, Altimeter, Striker, nvidia, and individuals including Jeff Dean, Brian Zhan, and Max; testingcatalog reports the $55M seed at a $300M valuation. A reposted claim that the team includes Apple's LLM lead, an xAI lead, and a Harvard professor appears only in secondhand accounts.
• Reception: Jeff Dean and Milanfar reshared praise; researcher cneuralnetwork noted their 1.5-year-old paper found early signs of visual reasoning in Gemini, now realized in Glimpse; a user's chessboard test scored Glimpse 8/10 vs. Gemini Flash's 5/10; svpino demoed arbitrary image/video QA such as tracking actions over time; lmoroney described the reasoning as "pinch-to-zoom while thinking."

Unconfirmed

• All benchmark results are company-reported; training cost, price multiples (6%, 17x), and funding figures lack third-party verification.
• Posts by dustinvtran carry a possible meme tone; his cited numbers (5 of 12 benchmarks at or above Astra; beating Gemini 3.8 Flash on real visual tasks) and claimed team membership should be treated cautiously, as should reposts claiming across-the-board superiority over Flash-class models.

Why it matters

If verified, Glimpse would challenge the consensus that frontier capability requires massive compute, approaching frontier vision reasoning at a tiny fraction of cost; its visual scratch space offers a new paradigm for multimodal reasoning. Key follow-ups: independent benchmarks and third-party replication on real dynamic visual tasks.

• Elorian Launches Glimpse, a 'Visual Thinking' Model Claiming Frontier-Level Benchmarks — AndrewDai · 2026-10-17
• Small team claims frontier multimodal model at 17x lower cost with hundreds of GPUs — dustinvtran · 2026-10-17
• Elorian launches Glimpse, a visual reasoning model claiming parity with GPT-6 Astra — AndrewDai · 2026-10-17
• Glimpse in detail: visual-space reasoning, multimodal tools, and subagents, free to try — AndrewDai · 2026-10-17
• AndrewDai launches Glimpse, a 'visual thinking model' claiming frontier-level visual reasoning vs Astra — AndrewDai · 2026-10-17
• Elorian launches Glimpse, a 'visual thinking' model claiming parity with GPT-6 Astra — AndrewDai · 2026-10-17
• Glimpse team thanks backers including Nvidia, Menlo Ventures and Jeff Dean — AndrewDai · 2026-10-17
• Glimpse model claims frontier multimodal performance at 17x lower token prices — dustinvtran · 2026-10-17
• Elorian's Glimpse claims frontier visual reasoning at 16x lower cost than Astra — thegautamkamath · 2026-10-17
• Elorian launches Glimpse, a visual thinking model for motion and viewpoints — dustinvtran · 2026-10-17
• Glimpse, first model from ex-Gemini team, beats Gemini 3.8 Flash 8/10 vs 5/10 on board test — dair_ai · 2026-10-17
• Jeff Dean backs Elorian's Glimpse, a visual thinking model competitive with Astra on key benchmarks — JeffDean · 2026-10-17
• Glimpse Launches, Bringing Visual Reasoning Research From Gemini Paper to Life — cneuralnetwork · 2026-10-17
• Ex-DeepMind team launches Glimpse, claims to beat GPT-6 Astra at 6% cost — damianplayer · 2026-10-17
• Ex-Google/Apple team launches Glimpse, a frontier-level visual thinking model on a few hundred GPUs — cihangxie · 2026-10-17
• Nvidia-backed Elorian launches visual reasoning model Glimpse, claims it beats GPT-6 Astra at 1/17 the price — rickasaurus · 2026-10-17
• Glimpse launches as a 'visual thinking model' claiming Astra-level reasoning benchmarks — ytjessie_ · 2026-10-17
• Elorian, founded by Gemini and Apple MM-1 researchers, launches visual reasoning model Glimpse — minchoi · 2026-10-17
• Glimpse launches on Elorian: visual intuition model scores 46.4% vs humans' 93.1% — testingcatalog · 2026-10-17
• Glimpse claims frontier multimodal performance at 17x lower token prices than rivals — dustinvtran · 2026-10-17

25 more related posts →

Episode 2 · Glimpse Visual Reasoning Model Launches with Astra-Level Capability at 1/17th the Cost (2026-10-20, 2 posts)

Andrew Dai and the Elorian team released Glimpse, a visual reasoning model claimed to match Astra-level performance on key public benchmarks while costing 17x less, targeting large-scale video analysis.

• Glimpse launches as a visual thinking model claiming Astra-level reasoning benchmarks — LinusEkenstam · 2026-10-20
• Elorian's Glimpse claims Astra-level visual reasoning at 17x lower cost — FinanceYF5 · 2026-10-20