LongCat-2.5-Preview vs Qwen3.8-Max: A Seventh of the Price and None of the Scoreboard

Meituan put LongCat-2.5-Preview on its API platform on 25 September 2026 with a changelog entry, a rate card, and no benchmark of any kind — then priced it at roughly a seventh of what Qwen3.8-Max charges for the same call. That is not a small gap, and it is not a safe one. Qwen3.8-Max, the vendor's flagship, has been generally available since 3 August 2026, carries an independent Artificial Analysis rank, publishes a dated snapshot you can pin by name, and bills 6.7× as much for input and 5× as much for output. The question this comparison has to answer is not which model is better — nobody outside Meituan is in a position to answer that yet. It is what you are buying with the difference, and whether the difference is worth buying.

What each side has actually published

Start with provenance, because it is the thing that separates these two more sharply than any benchmark would.

LongCat-2.5-Preview arrived with a dated entry on Meituan's own changelog — the vendor's words: "Version: 2026-09-25 — LongCat-2.5-Preview Now Available" — a pricing page listing it as the platform's current pay-as-you-go model, and a quick-start guide covering both wire formats. Meituan's own account on X describes it as "1.6T parameters. ~48B active. A 1M-token context window. Natively multimodal." Beyond that, there is nothing to read. No repository in the meituan-longcat Hugging Face organisation covers this model — the org's most recent repository is LongCat-Flash-Lite-Sparse, from 31 July — and the OpenCode model catalogue, which serves it, records it as closed-weight. No model card, no technical report, no licence, no parameter table published by the vendor, and no independent evaluation anywhere: as of 27 September there is no LongCat-2.5-Preview page on Artificial Analysis.

Qwen3.8-Max is the opposite case in almost every respect. It went generally available on 3 August 2026 with a published rate card, and Alibaba shipped a named refresh, Qwen3.8-Max-0902, on 2 September. Artificial Analysis measures it: Intelligence Index 45.4, ranked 15th of 145 models, and a Coding Index of 76.2, ranked 9th of 138. It posts GPQA Diamond 92.8%, Humanity's Last Exam 43.1%, SciCode 52.1%, Long-Context Recall 80.3%, Terminal-Bench 2.1 at 88.8%, and τ-bench banking at 47.8%. That is a measured model with a receipt for every number.

So the honest summary of the evidence column is lopsided in a way that has nothing to do with capability. One of these models can be evaluated before you commit. The other can only be tried.

The rate card, line by line

Both are million-token models with the same output ceiling, so the spec sheets converge exactly where a spec sheet is least useful:

• Uncached input — LongCat-2.5-Preview $0.30 per 1M vs Qwen3.8-Max $2.00 per 1M, a 6.7× gap.

• Cached input — $0.006 per 1M vs $0.25 per 1M, a 41.7× gap.

• Output — $1.20 per 1M vs $6.00 per 1M, a 5× gap.

• Context window — 1,048,576 tokens vs 1,000,000 tokens. Functionally a tie.

• Maximum output — 131,072 tokens on both. Identical.

• Independent score — none published vs AA Intelligence 45.4 and AA Coding 76.2.

Every LongCat figure there comes from Meituan's own pricing page, and the page labels the whole column "Discounted Price (limited-time)". Qwen3.8-Max's figures are Alibaba's list rate, read off our own model card, with a cache read at $0.25 and a cache write at $2.50 per million. Artificial Analysis's snapshot is dated 2 September 2026.

The cached-input line is the one to stare at. A 42× difference on cache reads is the largest single number in this comparison, and it lands on the dimension agent workloads actually live on. An agent loop that re-sends a 100,000-token system prompt and tool schema on every turn is paying the cached rate for most of its tokens, not the headline input rate.

A 150,000-token call, priced both ways

Take a plausible agent-shaped request: 150,000 input tokens, 50,000 of them served from cache, and a 4,000-token answer. On Meituan's discounted rate that call costs $0.0501. On Qwen3.8-Max's list rate, the identical call costs $0.3365 — 6.7× more, or $50 against $337 for a thousand calls a day. Push the mix fully cached, which is the steady state for a repeated prompt, and the ratio widens to 12.3×: $0.006 against $0.074 per call.

Nothing in that arithmetic depends on LongCat-2.5-Preview being good. That is precisely the problem. The multiplier you are being offered is real, and the word attached to it on the vendor's own page is "limited-time", with no end date and no stated successor price. A 6.7× discount with no published expiry is a budget line you cannot put in a plan; it is a budget line you watch.

There is also a routing asymmetry worth stating plainly. Qwen3.8-Max is a route we serve — qwen/qwen3.8-max and the pinned qwen/qwen3.8-max-0902 — at Alibaba's list price with zero markup, so a vendor price move lands on our side the same day rather than weeks later. LongCat-2.5-Preview is not one of our routes, and we are not going to pretend otherwise; on the cheap side of this comparison you are dealing with Meituan's own API and whatever platform you reach it through.

The one claim Meituan's own API contradicts

This is the finding that should decide the matchup for anyone whose workload is not pure text.

Meituan's changelog lists image understanding as the headline addition to the 2.5 generation: "Multimodal Understanding: New image understanding capability, able to parse image content, supporting cross-modal Q&A, content summarization, and complex visual reasoning." The X post says "natively multimodal." Those are vendor claims, and on their own they would be reasonable to carry into a decision.

Then the same documentation site publishes the example response for retrieving that model's own metadata, and the model it returns is text-only. Against the id "LongCat-2.5-Preview", the payload shows context_length: 1048576, input_modalities: ["text"], output_modalities: ["text"], and a modality string of text->text. No image entry. Not in an old snapshot — in the worked example on the page that documents the endpoint.

That does not prove the model cannot see. It proves the vendor has not reconciled its prose with its own API schema, and it means a buyer cannot bill a vision workload against that changelog sentence until someone settles it. Put it next to the other side: our catalogue lists Qwen3.8-Max accepting text, image and video input, with vision among its declared capabilities, and there is an independent benchmark table behind the model. If your task is document understanding, screenshot triage, or video-frame analysis, the expensive column is the one with a verified input modality and the cheap column is the one with an unresolved one. That single row can erase the entire 6.7×.

What a Preview cannot be pinned to

Alibaba ships dated snapshots. qwen/qwen3.8-max-0902 is a named, frozen build at the same $2/$6 as the base model, which means a routing rule that names it keeps returning the same weights next month. Behaviour changes arrive as a new id you can adopt on your own schedule.

LongCat-2.5-Preview has no such handle. The model id is the preview id, there is no snapshot suffix, no version history on the platform, and no changelog entry that would tell you the weights moved — only that the discount is "limited-time". It is a preview in the ordinary sense: the thing under the name is allowed to change, and you would find out from your output rather than from a release note.

The cheapest way to buy the missing evidence is the window Meituan opened on the other side of this: OpenCode lists LongCat-2.5-Preview as free for a limited period, with the provider following a zero-retention policy and not training on your data, and it said on 26 September that the window runs two weeks. Free input, free output, free cache reads. That is a legitimate way to build the benchmark table nobody has published — run your own evaluation corpus against it, and you have a number where the vendor gave you a sentence. What it is not is a price you can plan around: two weeks ends, and the number on the other side of it is Meituan's to set.

Who should pick which

Reach for Qwen3.8-Max when the workload is the reason you are buying the model in the first place. If the input is images or video, if the answer has to be reproducible build to build, if an independent index is a procurement requirement, or if the task is the kind of agentic coding that Terminal-Bench and the Coding Index are proxies for, the 6.7× is the price of being able to check your work. On one key it also sits behind a fallback chain, so a scored model can absorb a bad day for the unscored one without you running two integrations.

Reach for LongCat-2.5-Preview when the workload is high-volume, short-context, text-only, and internal — classification, extraction, summarisation, bulk rewriting — and when the cost of being wrong is a retry rather than a customer. On that shape the cached-input line is not a discount, it is the whole economics of the job, and 42× is a number worth the effort of measuring. Use the free window to produce the benchmark that does not exist. Do not migrate a critical path onto it.

What would change this verdict, in order of weight: an independent evaluation of LongCat-2.5-Preview; a published end date and successor price for the discount; a version history with dated snapshots; and a resolution of whether the modality list is text-only or not. Any one of those makes the cheap column more than a discount. Until at least one lands, this comparison is not a matchup between two models — it is a matchup between a price you can verify and a capability you cannot.