Glimpse vs Qwen 3.8: A Million-Token Window and a Number Nobody Has Published

One of these models publishes a one-million-token context window and a rate card you can budget against. The other publishes no context window, no parameter count and no architecture at all. Glimpse, released by Elorian on 16 October 2026, is a research preview whose entire published capability is a behaviour — it can return to a specific moment in a video while it is still thinking, re-scan a nine-tenths-of-a-second window at a denser frame rate, zoom into a region, or follow one feature across time before it commits to an answer. Qwen 3.8 is A​libaba's August 2026 generation, and the member this comparison actually turns on is Qwen3.8-Max: live as a priced API since 3 August 2026 at $2.00 per million input tokens and $6.00 per million output, with a million-token window and native image input. Qwen3.8-Max is one of the models we route; Glimpse is not on our catalogue at all. That asymmetry is not a small procedural detail. It is the shape of the decision, because one of these is a call you can put in a budget line and the other is a preview you have to try before you can cost it.

What each one is actually built to do

These are not two answers to one question, and the tell is that their bottlenecks sit in different places.

Qwen3.8-Max's bottleneck is capacity and cost at scale. A million tokens of context means an hour-long recording, a thousand-page filing, or a week of logs goes into a single call without a chunking pipeline in front of it. The model is natively multimodal on input — text and images in, text out — so a document-heavy or screenshot-heavy workflow does not need a separate vision stage. Its parameter count has never been disclosed by A​libaba, but the open sibling of its core, Qwen3.8-2.4T-A95B, is a 2.4-trillion-parameter mixture-of-experts with roughly 95 billion active, which is as close as anyone outside the company gets.

Glimpse's bottleneck is resolution in time. Its window is not the constraint; the sampling schedule is. Ordinary video understanding models sample frames on a fixed grid, and whatever falls between two sampled frames is gone — which is why "did the camera pan or cut" and "is that the same object after it was occluded" are hard questions rather than easy ones. Glimpse addresses that by spending extra compute on one interval after it has already seen the whole clip. In the trace Elorian published, it takes a first pass at roughly 8 frames per second, forms a tentative reading, then returns to 0.9 seconds of the clip at around 14 frames per second to settle whether it is looking at a pan or a cut.

• Context — 1,000,000 tokens for Qwen3.8-Max, published by A​libaba and listed on our own route. Not published for Glimpse, which is a gap a reader should treat as a gap rather than an omission to be filled with an estimate.

• Input modality — both accept images. Qwen3.8-Max accepts video frames on a schedule you choose. Glimpse accepts video and can revisit an interval after a first pass, which is the whole product.

• Size — undisclosed for Glimpse. For Qwen3.8-Max, undisclosed as a served model, with a 2.4T / ~95B-active open sibling as the closest available proxy.

• Weights — none for Glimpse; it is API and chat only, with a free sign-up. Mixed for the generation: Qwen3.8-27B is a dense 27-billion-parameter model under Apache 2.0, the 2.4T core is downloadable under a custom licence, and Qwen3.8-Max itself has no weights at all.

• Price — Glimpse $0.60 / $3.00 per million tokens. Qwen3.8-Max $2.00 / $6.00, with a $0.25 cached read.

• Independent score — Artificial Analysis Intelligence Index v4.3.2 puts Qwen3.8-Max (0902) at 45, ranked 33rd of 217, and measures its output speed at 38.8 tokens per second, which AA describes as notably slow for its tier. Glimpse has no independent score of any kind; it shipped yesterday.

Both vendors chose their own comparison set

Neither model's headline pricing claim survives contact with the other's.

Elorian frames Glimpse at $0.60/$3.00 as 94% below GPT-6 Astra at $10/$50, 85% below Claude Opus 5.5 at $4/$20 and 70% below G​PT-6.1 Sol at $2/$10. Every one of those is a frontier flagship. Set against them, Glimpse looks like the cheap option, and the claim is arithmetically correct.

Set Glimpse against the cheap end of the Qwen 3.8 line and the framing inverts. Qwen3.8-Flash runs at $0.15 per million input tokens and $0.47 per million output, with cache reads at $0.018 and cache writes at $0.230 — so on a workload with a large stable prefix, the effective input rate is a fraction of a cent per million, an order of magnitude under Glimpse's input price and roughly two orders below its output price. Qwen3.8-Max, the flagship, is more expensive than Glimpse on tokens and still four times cheaper than the frontier model Elorian priced against.

That is not a contradiction on either side. It is two companies choosing the comparison that flatters them, and the reader's job is to replace both with the workload in front of them.

Pricing an hour of video, both ways

Abstract rate cards decide nothing, so put a concrete job through them: a surveillance-style review task, one hour of footage per case, 200 cases a month, a question that needs the model to answer in prose with timestamps.

On Qwen3.8-Max the hour fits in the context window without chunking. At a rough 800,000 input tokens for a densely sampled hour plus a 1,500-token answer per case, a month costs about $322 in input and about $1.80 in output — call it $324, with the caveat that if the same footage is reviewed repeatedly, the $0.25 cached read cuts the input line hard and this arithmetic stops being the right one.

On Glimpse the hour also fits, and the mechanism is aimed at exactly the question being asked. Elorian publishes no per-hour figure, but it does give one task-level number — a robot-planning episode costing $0.035 against $0.249 on Astra — and there is no honest way to extend that to video review without inventing tokens. What can be said is that the pricing shape is different in kind: Glimpse charges per token on a model that deliberately spends extra tokens looking again, so the rewind is not free and the trace is the receipt. Whether that is $300 a month or $900 depends on how often the model decides to go back, and that number only exists after you run your own clips.

The other half of that bill is worth stating plainly, because it is the part that does not appear on a rate card. Qwen3.8-Max is one of our routes, at the provider's list price passed through with nothing added per token — so if A​libaba moves that rate, the change is live on our side the same day rather than at the next contract renewal. Glimpse is not on our catalogue; Elorian's own API and chat interface are where you get it, and that is a statement about what we host rather than a claim about the model. Where a routing layer proves useful in this particular comparison is one layer up: one key in front of 200-plus models, automatic failover, so an unproven preview or a slow flagship sits on a side path rather than in front of your production route.

Two other Qwen 3.8 members change the answer

Comparing Glimpse to Qwen3.8-Max alone gives a misleadingly simple picture, because the generation contains members that behave nothing like its flagship.

Qwen3.8-Flash is the volume lane: $0.15/$0.47 for the same short task run a million times rather than one large call. For a workflow that already samples frames on a fixed schedule and just needs a cheap reading of each one, Flash is the correct instrument and the Glimpse question never arises.

Qwen3.8-27B is the member that changes the wiring diagram rather than the price. It is a dense 27-billion-parameter multimodal model under Apache 2.0 that you can download and run on one machine, which means the footage does not leave your infrastructure at all. Qwen3.8-Max at $2/$6 answers what a video says; the 27B answers whether you are allowed to send the video anywhere in the first place. Neither Glimpse nor Qwen3.8-Max offers that, and it is not a capability either vendor is likely to add.

So the practical map has three lanes, not two. Cheap fixed-schedule sampling at Flash rates; large-context multimodal reasoning with Max; and a self-hosted 27B when custody of the media is the binding constraint. Glimpse is not a fourth lane — it is a bet on the one question none of the three can answer well, which is what happened inside a sub-second interval.

What would change this comparison

This page is worth re-reading when three specific things happen, and none of them has happened yet.

The first is any independent evaluation of Glimpse. Its benchmark table is vendor-reported, unreproduced, and run at the model's own highest thinking effort with its native agent harness, with a footnote disclosing that several competitors were run below their maximum. One independent harness reproducing a single row would convert a claim into evidence.

The second is Glimpse's context window. A video-reasoning feature whose window is smaller than the footage a real workflow hands it is a demo, and the number has not been published.

The third is a Qwen3.8-Max rate change or a new member in the generation. The line has moved once already — Max entered at $2/$6, below the prior generation's preview pricing — and AA has evaluated the Max line at more than one build with the index number moving between them. A generation that ships five members in a quarter does not stop at five.

Until then the decision rule is narrow enough to state in one line: if your question is what happened inside a fraction of a second of footage, Glimpse is the only one of the two whose stated mechanism aims at it, and the free sign-up makes the trial cheap — run a clip, then read the visual scratch space and check whether the model ever went back to the interval that decided the answer. If your question is what a long document, a screenshot set or an hour of footage says as a whole, and the answer has to survive a budget review, Qwen3.8-Max is the model with a window, a rate card, an independent index and a route that prices it, and Glimpse was never in that race.

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.