Elorian Glimpse: the model that zooms, replays, and codes while reasoning about an image or a video

Elorian Glimpse: the model that zooms, replays, and codes while it reasons over an image or a video

🔎 An unknown from Palo Alto takes on the vision giants — at 6% of the price

On October 16, 2026, Elorian, a Palo Alto startup that was invisible the day before, released Glimpse: a "visual thinking model" claiming frontier-level performance, "sometimes surpassing GPT-6 Astra and Claude Opus 5.5," for $0.60 per million input tokens and $3 for output. Elorian's official announcement describes a model that reasons about the physical world based on what it sees.

The thesis is not to build a bigger model, but to change the way it looks. Traditional VLMs tokenize the image only once, then reason in text over a fixed encoding. Glimpse maintains a "visual scratch space": it zooms in on a region, replays a video at a higher frame rate, runs code to take measurements — while it responds.

The timing is no accident. Andrew Dai, founder and ex-Google DeepMind, explained as early as July 2026 that progress has been "extremely uneven in visual understanding and visual reasoning." With $55M raised before any product and only a few hundred GPUs, Elorian is betting that looking better beats spending more.

The essentials

• Glimpse (Elorian, October 16, 2026): a model that returns to the image during its reasoning — zooming, frame extraction, code execution — and can spawn and coordinate subagents exchanging text and visual artifacts.
• Aggressive pricing: $0.60/M input tokens, $3/M output tokens (October 2026), i.e. 94% cheaper than GPT-6 Astra ($10/$50) and 85% cheaper than Claude Opus 5.5 ($4/$20).
• Self-reported benchmarks: ahead on scientific chart reasoning and short-form video captioning, behind on raw perception (BabyVision 95.36 vs 97.68 for GPT-6 Astra).
• Already-funded startup: $55M seed at a $300M valuation before any product (July 2026), with Nvidia and Jeff Dean in the round.
• Access: API open in research preview, chat on waitlist; neither license nor context window announced to date.

Recommended Tools

Glimpse is entering an already structured market. Here's where it stands against the multimodal models it directly targets, with pricing announced in October 2026.

Prices per million tokens, based on vendor announcements (October 2026 — check official websites before committing).

The "visual scratch space": looking again, instead of guessing

The real novelty of Glimpse isn't a benchmark score, it's a cognitive loop: the model can return to the visual representation during its reasoning, instead of working from a single, definitive tokenization.

Concretely, the "visual scratch space" allows it to crop a suspicious area, extract frames from a video, run code to measure or compare two regions — all while it builds its answer. This is the mechanism TechCrunch describes at launch.

The model was trained end-to-end to use multimodal tools: writing code, transforming images, watching videos. It can also spawn and coordinate subagents that exchange text, visual artifacts, and internal representations, the official announcement specifies.

Subagents that share images, not just text

Most agent frameworks multiply text exchanges. Glimpse circulates visual representations between its subagents: one zooms in, another measures, a third arbitrates. It's a rare architecture — and Elorian remains discreet about its real-world production performance.

Writing code to think isn't new: it's the daily routine of code agents, as shown by our selection of the best AI tools for code. The trend is even accelerating — JetBrains just released Mellum2.1, an open "thinking" 12B model dedicated to code agent infrastructure. Glimpse's novelty lies in the object of that code: interrogating the physical world rather than producing a program.

Andrew Dai showed TechCrunch a photo of a game board that no frontier model could read. By zooming in square by square during its reasoning, Glimpse manages it. The maze demo is even more striking: 5/5 solved, versus 1/5 for Gemini 3.8 Flash — figures from Elorian, relayed by daily.dev.

My take: it's the same shift as elsewhere in AI — reasoning instead of searching by similarity, the philosophy championed by PageIndex, the vector-free RAG that reasons instead of searching. Here, we reason instead of guessing from a single encoding.

Benchmarks: frontier on in-house evals, lagging on perception

Elorian's numbers are real, but self-reported — and the official table itself shows setbacks against GPT-6 Astra. Read it before the press releases.

The announcement claims frontier-level performance "sometimes surpassing" GPT-6 Astra and Claude Opus 5.5. On the two tasks where Glimpse excels — scientific chart reading and short-form video captioning — TechCrunch confirms that in-house evaluations place the model ahead of Astra.

But a recount published on dev.to re-tallied the results: against GPT-6 Astra, Glimpse wins 4 rows, ties 1, and loses 6 out of 11. The two biggest gaps — HSS and Video-MME v2, where GPT-6.1 Sol leads at 65.27 — are around 7 points. It's neither a collapse nor an overtaking.

The community saw the problem clearly: 30% positive reactions, 45% mixed, 25% skeptical on daily.dev, essentially because the benchmarks are self-reported. The specialist press outlets — MarkTechPost, MadeWithModels, ToolSift — covered the launch without independent verification at this stage.

My verdict: wait for third-party evaluations before rewriting your stack. In the meantime, judge on your own use cases — it's measurable in an hour with the API.

The price: 94% less per token, not necessarily per bill

At $0.60/M for input and $3/M for output, Glimpse is the most aggressive offering in the mid-level segment. But a bill is measured in tokens consumed, not in the advertised rate — and Glimpse looks several times.

Prices announced in October 2026 — check the vendors' websites. TechCrunch positions Glimpse as "cheaper than mid-level models like Sonnet and Sol, but more expensive than lightweight models like Haiku and Luna." The message is clear: Elorian is targeting the mid-market with frontier ambitions.

Elorian's marketing figure: planning the LIBERO robotic arm costs $0.035 with Glimpse versus $0.249 with GPT-6 Astra — roughly 7 times less per task (daily.dev). But the dev.to analysis warns: the post only gives a single per-task figure, and a token price 94% lower only becomes a 94% lower bill if both models spend the same number of tokens.

Yet Glimpse's very mechanism — repeated zooms, frame extraction, code execution — can inflate consumption. It's the classic trap of "reasoning" models. My advice: measure the cost per task on your own data, never the per-token rate.

Andrew Dai, $55M and a few hundred GPUs: Elorian's thesis

Elorian isn't just another lab. It's a targeted bet on visual reasoning, led by someone who drove Gemini's data.

Andrew Dai spent 12 years at Google, led teams at Google DeepMind and drove Gemini's data, notes Dealroom News. His team includes members who led Gemini at Google and MM-1 at Apple (daily.dev).

The funding is aggressive: a $55M seed at a $300M valuation raised before any product, in July 2026 — a valuation-to-capital ratio tighter than Thinking Machines', notes TechCrunch. The round was co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with Nvidia and Jeff Dean, Google's chief scientist, among the participants.

His quote at the time: "progress has been extremely uneven in visual understanding and visual reasoning." That's exactly the niche Glimpse is attacking — neither perception nor generation, but the loop between the two.

A telling detail: post-training was done on an open-weights base, with only a few hundred GPUs. At a time when frontier labs are stacking tens of thousands of chips, Elorian is betting on efficiency — a line also followed by the open-weight ecosystem, like Flux 3 Image Klein, the 56B open-weight image model that runs on a single card.

The stated goal, reported by AI In Context: the "best visual reasoning model in the world" next year. Ambitious for a team that just shipped its first product.

Use cases: where Glimpse makes sense — and where it doesn't

Glimpse is worth a look when the answer is hidden in the pixels: charts, counting, video events, robotic manipulation. It loses its appeal when raw perception or volume is the priority.

The claimed gains, as summarized by Dealroom: scientific chart reasoning, counting, short-form video captioning, understanding of robotic manipulation events. Four task families, one common denominator: you have to look twice.

Robotics: the LIBERO cost ($0.035 per planning) paves the way for cheap vision as a planner. This is a structural trend — Stanford has shown with DexAgent that a single human video is now enough to train robotic hands. Understanding manipulation events 7 times cheaper changes the economics for robotics labs.

Short-form video: short-form captioning is a stated strength. Combined with streaming VLMs that anticipate what comes next in a sequence — see Foresight, the streaming VLMs that can already anticipate the future — video analysis becomes a real product workstream in 2026. And on the generation side, Gemini Omni pushes any-to-any all the way to video output: the market is specializing at each end of the pipeline.

Charts and scientific data: reading a graph requires measuring, comparing, zooming in. Exactly what the scratch space does, with code executed mid-reasoning.

Three cases where I'd advise against Glimpse today:

• Maximum raw perception: GPT-6 Astra still leads on BabyVision and HSS.
• OCR and high-volume image sorting: lightweight models like Haiku or Luna remain cheaper per token.
• Anything that requires a known context window: it isn't declared.

How to test Glimpse today

The API is open right away in research preview; chat, on the other hand, remains on a waitlist. Here's how to evaluate the model without getting burned.

1. Get API access on Elorian's website — the research preview is open, as confirmed by the official announcement and AI In Context.
2. Pick three tasks with verifiable answers: chart reading, object counting, a specific question about a video. That's where the scratch space flexes its muscles.
3. Measure cost per task, not per token: compare against your current model on identical inputs, total consumption included.
4. Check the license and context window before any production deployment — neither has been disclosed so far (dev.to).

To build a small web demo that calls the API — a quality control dashboard, chart data extraction, image sorting — a VPS from Hostinger is more than enough to get started (pricing: check hostinger.com, October 2026).

And if your pipeline mixes vision and code generation — Glimpse writes code during its reasoning — our comparison of the best LLMs for coding will help you pick the text component. On the open-source side, Kimi K2.7-Code now cuts 30% of reasoning tokens and beats Opus at tool use: frugal agents are everywhere.

One last point: this is a research preview. Pricing, quotas, and availability can change; Hyper.ai confirms the current rate, but double-check before committing to any volume.

What Glimpse Doesn't Say

The dossier is incomplete, and you should know that before investing engineering time.

• Context window: no value declared in the launch post (dev.to). For long-form video, that's a blocker.
• License: nothing announced. Post-training on open-weights bases says nothing about the terms of use of the final model.
• Subagent architecture: Elorian describes the capability, not the costs or coordination limits.
• Chat availability: waitlist, with no announced date. Only the API is open.

None of this is suspicious for a research preview — it's normal. But it does mean one thing: don't build contractual commitments on a model whose terms can change overnight.

❌ Common Mistakes

Mistake 1: confusing price per token with price per task

The $0.60/$3 pricing applies per token. If Glimpse spends three times as many tokens because it looks three times, the advantage melts away. Solution: benchmark the full cost per task on your real data, as the dev.to analysis recommends.

Mistake 2: mistaking self-reported benchmarks for independent results

The numbers come from Elorian, based on in-house evaluations. The dev.to review shows 6 losses out of 11 against GPT-6 Astra. Solution: wait for third-party evaluations, or test it yourself on your own cases before deciding.

Mistake 3: deploying a preview to production without a safety net

Research preview, undisclosed license, unknown context window, chat behind a waitlist: the contract isn't complete. Solution: prototype, but don't migrate your critical workflows until there's a stable release and clear terms of use.

Mistake 4: using Glimpse where raw perception is what matters

On BabyVision and HSS, GPT-6 Astra still comes out ahead. If your bottleneck is first-pass perception quality — medical imaging, high-resolution quality control — a conventional frontier VLM remains the right choice. Glimpse shines when it can take another look, not when it has to get it right on the first pass.

❓ Frequently Asked Questions

Is Glimpse open-source?

No. The post-training is built on an open-weights base, but Elorian does not publish Glimpse's weights and declares no license (October 2026). Access goes through the API in research preview; the chat is on a waitlist. Watch Elorian's blog for any potential changes.

How much does Glimpse cost?

$0.60 per million input tokens and $3 for output (October 2026, Elorian announcement, confirmed by Hyper.ai). That's 94% less than GPT-6 Astra ($10/$50) and 85% less than Claude Opus 5.5 ($4/$20). Note: the actual bill depends on the number of tokens consumed per task.

How do I access Glimpse?

Two doors: the API, open immediately in research preview, and a chat interface accessible via waitlist. The official announcement and AI In Context confirm this dual availability. Check quotas and service stability before committing to any volume — it's a preview, not a general-purpose product.

Does Glimpse replace GPT-6 Astra or Claude Opus 5.5?

No, not yet. It loses 6 of 11 comparisons against GPT-6 Astra according to dev.to, including BabyVision (95.36 vs 97.68) and HSS (46.36 vs 53.6). It wins on scientific chart reasoning and short video captioning. Today, it's a targeted and cheap complement, not a successor.

Who is behind Elorian?

Andrew Dai: 12 years at Google, former director at Google DeepMind, where he led Gemini's data. The team includes members who led Gemini (Google) and MM-1 (Apple). Elorian raised $55M in seed at a $300M valuation in July 2026, a round co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with Nvidia and Jeff Dean participating.

✅ Conclusion

Glimpse doesn't dethrone the perception giants, but it makes "looking again" a viable and cheap reasoning strategy — and that's a first. If your workflows hide answers in pixels, test the API and measure the cost per task; for the code component of your pipeline, rely on our comparison of the best LLMs for coding.