Elorian's Glimpse Rewinds the Video Before It Answers

The most interesting sentence in Elorian's launch post for Glimpse is not about a benchmark. It is about a 0.9-second window. In one of the worked traces the company published on 16 October 2026, Glimpse runs a first pass over a clip at roughly 8 frames per second, reaches a tentative reading, then goes back and re-scans that nine-tenths of a second at around 14 frames per second to decide whether it is looking at a camera pan or a cut. A second trace, in the same post, follows a single window edge from about 2.2 seconds into a clip until it comes to a near stop at 5.13 seconds. When a preview model sits in a production path, the failure modes that matter are the ones already logged in the incident archive, which is where we track the ones that have actually happened.

That is the pitch, and it is a genuinely different one. Glimpse is Elorian's first model: a visual-reasoning system that the company describes as reasoning in visual space rather than converting an image into language and reasoning there. Elorian itself is not obscure — it came out of stealth in April 2026 behind Andrew Dai, a former G​oogle researcher, with $55 million from Striker Ventures, Menlo Ventures and Altimeter and participation from N​VIDIA, 49 Palms and Jeff Dean, and Forbes put it on its Next Billion-Dollar Startups list in July. Glimpse is a research preview, available now through an API and a chat interface, and it is the first thing the company has actually shipped.

What "rewinding" means, mechanically

Elorian calls the display surface the visual scratch space. The company's description of what the model does inside it is specific: it may zoom into a region, watch a short interval at a denser frame rate, or track one feature across time. The post frames the behaviour as "questioning its first impression, and returns to the visual evidence before it answers."

Two properties of that are worth separating, because they get conflated.

The first is re-entrant sampling. Ordinary video understanding models sample frames on a fixed schedule — often a uniform grid, sometimes a keyframe heuristic — and whatever falls between two sampled frames is gone. Glimpse can spend extra compute on a specific interval after it has already seen the whole clip. The 8-to-14 fps step-up is that capability in its smallest form. It matters most on exactly the questions fixed sampling fails: pan-or-cut, whether two appearances of an object are the same object, what happened in the half-second around an event.

The second is the trace itself. Elorian's platform shows the model's scratch space as it reasons, which means the evidence for an answer includes where the model looked. That is a different kind of audit trail from a chain of text tokens, and it is the part with obvious downstream value — an answer about a construction site that comes with the timestamps it relied on is reviewable in a way a paragraph of prose is not.

Underneath both, the company says the model was trained end to end to use multimodal tools — writing code, transforming images, watching video — and can spawn and coordinate subagents that exchange text, visual artifacts and internal representations. It was, in Elorian's words, post-trained on an open-weights base model by a small team with "only a few hundred GPUs," and is positioned as "a substantially smaller and cheaper model" than the frontier systems it benchmarks against.

The half of the benchmark table that matters

Glimpse's own results page publishes a table against GPT-6 Astra, and these are vendor-reported numbers with no independent reproduction. What is unusual is how mixed the table is. On the measuring tasks, Glimpse wins or ties. On the harder inference tasks, it loses.

• WGO Bench (manipulation events) — Glimpse 42.70 vs GPT-6 Astra 39.42.

• CharXiv-R — 92.90 vs 92.10.

• UGC-VideoCap — 82.66 vs 80.32.

• FSC-147 — 83.36 vs 82.02.

• ERQA — 83.75, level with Astra.

• E.T. Bench — 63.55, ahead of G​emini 3.8 Flash's 63.45.

• BabyVision — 95.36 vs Astra 97.68.

• DREAM-1K — 76.04 vs 79.05.

• TOMATO — 93.60 vs 96.02.

• Video-MME v2 — 56.68 vs Astra 64.06 and GPT-6.1 Sol 65.27.

• CountQA — 82.66 vs 86.42 for both Astra and Sol.

• HSS — 46.36 vs Astra 53.6.

There is a footnote attached to the table that a reader should carry into every one of those rows: the results use the highest thinking effort with the model's native agent harness, and for most video evaluations Sol and Opus 5.5 were run at high rather than max effort. The HSS row mixes in paper-reported numbers from other labs. None of this invalidates the table, but it means the comparison is Glimpse at its best-effort configuration against several competitors at something less than theirs — and Glimpse still loses the inference-heavy half.

The honest read of that split is narrow and useful. Re-look reasoning buys you better perception: counting, describing, reading charts, tracking a small object through motion. It does not buy you the thing HSS measures, which is what a scene implies. More looking is not more understanding, and the model's own table is the evidence.

What the price claim is actually claiming

Glimpse is priced at $0.60 per million input tokens and $3.00 per million output. Elorian spells out the comparison in its own post: 94% below GPT-6 Astra at $10/$50, 85% below C​laude Opus 5.5 at $4/$20, 70% below GPT-6.1 Sol at $2/$10, and 20% below G​emini 3.8 Flash's introductory rate of $0.75/$3.75. The post also gives a task-level figure — one robot-planning episode cost $0.035 on Glimpse against $0.249 on Astra.

Those are the vendor's numbers about the vendor's model and they are not independently checked. Two things about them are still worth noting. The first is that the percentage claims are all measured against frontier flagships, not against the cheap end of the market; on tokens alone Glimpse is squarely mid-market, and the article next to this one on Qwen 3.8 works through what happens when you put it beside A​libaba's volume tier instead. The second is that a rate card on a research preview can move at any time, and this is the one place where the routing layer earns its mention: OrcaRouter passes provider list price through with nothing added per token, so a vendor price change on a model we serve is live on our side the same day, and automatic failover means an unproven preview can sit behind a production path rather than in front of it. Glimpse is not on our catalogue today — Elorian's own API and chat interface are where you get it — so that is a statement about how we price, not a claim to host it.

The benchmark Glimpse lost on is its own

Nine days before the model shipped, Elorian published a benchmark. Humanity's Sixth Sense, built in partnership with Scale AI, measures implicit visual reasoning — what a scene suggests rather than what it states. It is 522 tasks, 288 still images and 234 short video clips, 17.6 hours of video, scored with human-written rubrics across four domains. Twenty-five models from eight companies were tested.

The spread is the story. Human baseline is 93.1%. The best published model comes in at 53.6. The median model scores 30.9%, and a blind control with the media stripped out still scores 6.6% — which tells you how much of the task is answerable without looking. Of the labelled failures, 94% trace to seeing or inferring rather than to reasoning: 53% misreading, 41% missed implications, roughly 5% logic errors. In one case, raising reasoning effort from high to xhigh cost a model 14 points.

Glimpse scores 46.36 on it. That is well above the median and above C​laude Opus 5.5's 44.6 and GPT-6.1 Sol's 46.6 — but behind GPT-6 Astra's 53.6, and on a benchmark the company built. Publishing a benchmark your own model fails to top is a credibility signal, and it also tells you precisely what the rewind does not fix.

What Elorian has not published

This is the section most launch coverage will skip, so it belongs here plainly. The Glimpse post gives no parameter count. No context window. No model size variants. No description of the architecture beyond "post-trained on an open-weights base model," which does not name the base model. No pretraining data. No output-token limits, no rate limits, and no API endpoint documentation beyond the platform itself.

That is not unusual for a research preview, and it is not a reason to distrust anything above. It is a reason to be careful with one specific kind of claim: any figure for Glimpse's size, context or throughput circulating this week is an estimate rather than a disclosure, and treating an estimate as a specification is how a capacity plan goes wrong.

The company is also explicit about what the model does not do yet. Robot action generation is listed as future work; calibrated reconstruction and robot control are described as "work ahead." On the worked tire-pressure example, the answer is called a qualitative prediction from a photograph with no trajectory or impact force simulated. And some of the examples are labelled as reviewed reference answers rather than recorded model traces, with caption and transition traces condensed for publication — a disclosure worth respecting rather than skimming.

Who should do something about this now

If you have video work where the failure mode is temporal rather than semantic — object identity across a cut, an event inside a sub-second window, a count that changes when something occludes — Glimpse is the first hosted model whose stated mechanism directly targets that, and the free signup makes the test cheap. The traces make the failure legible too: if the scratch space shows the model never re-scanned the interval that mattered, you have learned something about your pipeline and not just about the model.

If your need is inference about what a scene implies, wait. The vendor's own HSS table says the rewind does not get you there, and nothing in the release suggests the next version will close a 7-point gap in one step. Watch three things over the next two months: whether Elorian publishes the base model it built on, whether any independent harness reproduces a row of that table, and whether the context window turns out to be large enough to hold the video a real workflow hands it. All three are currently blank, and all three decide whether Glimpse is a capability or a demo.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.