Step 5 Preview vs LongCat 2.5: The Cheaper Model Is the One Nobody Can Score

On paper this is the easiest cost decision in the 2026 model market. LongCat 2.5 — the model Meituan actually serves under the identifier LongCat-2.5-Preview — prices uncached input at $0.30 per million tokens and output at $1.20, which is a third and less than half of what Step 5 Preview charges at $1.00 and $2.70. Meituan also gives you a 1M-token context window for that money. The catch is that nobody, including Meituan, has published a single benchmark number for it, while Step 5 Preview has one independent score, a downloadable checkpoint and a technical write-up. So the decision is not between a cheap model and an expensive one. It is between a model you can measure and a model you would have to measure yourself.

That is not a reason to dismiss LongCat 2.5. At ¥2 per million input tokens — roughly $0.30 on the listings that convert it — measuring it yourself is affordable in a way that measuring a frontier model is not. It is a reason to go in with the right expectations, and to know exactly which of the two vendors publishes what.

What each vendor will actually tell you

StepFun is forthcoming to the point of excess. Step 5 Preview went to its API on 20 September 2026 and to downloadable weights on 15 October: BF16 and FP8 on Hugging Face and ModelScope, vLLM and SGLang recipes on the same day, and a technical blog describing a 92-layer stack, sparse GQA attention with 3:1 interleaving of sliding-window and sparse full-attention layers, 8-of-352 experts active plus one shared, and a 600-billion-parameter model with 27 billion active per token. There is a rate card, a context limit, a licence with named revenue triggers, and one independent score.

One small thing worth noting about Meituan's own rate card: the column header reads "Discounted Price $/1M Tokens" and every figure underneath it is in yuan. The dollar numbers that circulate for this model are conversions somebody else did, not a second price list.

Meituan is forthcoming about the interface and silent about the model. The LongCat API Platform has an OpenAI-compatible endpoint and an Anthropic-compatible one, a published rate card, a published context limit, and documentation for billing, rate limiting and prepaid balance. What it does not have is a parameter count for LongCat 2.5, a technical report, a benchmark table, a Hugging Face repository, or a deprecation window. The platform's supported-models list contains exactly two entries — LongCat-2.5-Preview and LongCat-2.0 — and the 1.6-trillion-parameter, 48-billion-active figure that circulates attached to 2.5 is LongCat 2.0's model card, carried across without being re-published. This page will not print it as a LongCat 2.5 specification, and neither should anything else.

The price gap, line by line

• Uncached input per million tokens — Step 5 Preview: $1.00. LongCat 2.5: ¥2 on Meituan's own card, which third-party listings convert to about $0.30.

• Output per million tokens — Step 5 Preview: $2.70. LongCat 2.5: ¥8, converted by third-party listings to about $1.20.

• Cached input — Step 5 Preview: a 90% discount on the input rate, and ¥0.70 per million on the renminbi card after the cache-hit rate doubled from ¥0.35 at 00:30 Beijing time on 14 October 2026. LongCat 2.5: ¥0.04 per million, converted to about $0.006.

• Context window — Step 5 Preview: 1M tokens. LongCat 2.5: 1,048,576 tokens.

• Input modalities — Step 5 Preview: text, images and video. LongCat 2.5: documented as text in the API reference, described as image-capable in the release changelog — see below.

• Weights — Step 5 Preview: downloadable under StepFun Model License 1.0. LongCat 2.5: none. Meituan's Hugging Face organisation holds the LongCat 2.0 family, not 2.5.

• Independent score — Step 5 Preview: 29 on Artificial Analysis Intelligence Index v5.0. LongCat 2.5: no evaluator has published one, and no page exists to hold it.

• Stability commitment — Step 5 Preview: a published licence with named thresholds. LongCat 2.5: no deprecation window, no versioning promise, and a rate card whose own header says "limited-time".

One has vision. The other has a contradiction.

Step 5 Preview accepts text, images and video and returns text. That is unambiguous in the vendor's documentation and consistent everywhere in it.

LongCat 2.5 is described two different ways by two pages of Meituan's own documentation set. The changelog that announced the model on 25 September 2026 lists native image input among its features. The API reference, in the machine-readable field that SDKs and gateways actually read, lists input modalities as text only. One of those is wrong. It is a one-line fix to a generated page, and until it is fixed, the only safe assumption for a production integration is that image input is not available — a changelog entry is a promise and a schema field is a fact.

If image understanding is part of why you are shopping, that single contradiction ends the comparison. If it is not, it is a useful reminder of the general state of the documentation.

"Preview" means something different in each name

Both models carry the word, and in neither case is it decoration — but they carry it differently.

In Step 5 Preview, "Preview" is part of the product's identity as shipped: the string you send is step-5-preview, and StepFun's promotion path to a general-availability model is not something the documentation describes. What it does describe is a licence that will not change under you, and a checkpoint that will run tomorrow.

In LongCat-2.5-Preview, "preview" is load-bearing in three concrete ways. It is the only identifier the platform accepts — there is no GA alias to send instead. The price is explicitly provisional, with the words "limited-time" over every column and a closing sentence stating that the figures shown on the platform and in your billing records are authoritative. And there is no published stability commitment anywhere: no deprecation notice period, no versioning promise, and no statement about whether the model keeps its ID when it graduates.

That is normal for a preview and it is not a reason to avoid the model. It is a reason not to build a unit-economics model on a rate card that tells you not to.

Speed, where only one of them has been measured

Artificial Analysis measures Step 5 Preview at 179.1 output tokens per second on StepFun's API, against a class median of 74.1, at $2.35 per full index task. No equivalent measurement exists for LongCat 2.5, and the anecdotes that do exist point the other way: third-party hands-on write-ups of building front-end and animation demos with it report tasks taking upwards of thirty minutes end to end. That is not a token-per-second figure and it does not establish throughput — agent tasks spend most of their wall clock waiting on tool calls. But it does suggest that the model is not being marketed on speed, and a long-horizon loop on a slow model is a different product from a long-horizon loop on a fast one.

If your workload is interactive, this is the dimension to test first, before cost. A model at one third the price that takes three times as long per turn is not cheaper for a user waiting on it.

How to test a model nobody has scored

The practical answer is to stop comparing on paper and measure on your own workload, which is cheap here: at ¥2 per million input tokens, a thousand realistic tasks is a small bill. What you need is a way to run the comparison without building two integrations and reconciling two bills.

That is the part of this that OrcaRouter earns. Neither LongCat 2.5 nor Step 5 Preview is in our live catalogue — 214 model IDs across 16 providers, with no Meituan entry and no StepFun entry as of 29 October 2026 — so for these two specifically you are calling Meituan's platform and StepFun's API respectively. But the surrounding question, which candidates from the catalogue you should be benchmarking against, is one you answer under a single key: the same request shape against several models, one bill at provider list price with no markup, and automatic failover when you promote whichever one wins into production. The point of a catalogue that size is that the model you end up on is rarely the model you started evaluating.

Three questions this comparison raises, answered

Is LongCat 2.5 actually a 1.6-trillion-parameter model? Unknown. Meituan publishes no parameter count for 2.5. The 1.6T total and 48B active figures belong to LongCat 2.0's published model card, and the shape may or may not have been inherited. A vendor that will publish a rate card and a context limit but not a parameter count has made a choice, and the honest reading of that choice is that the number would not help the pitch.

Which one is cheaper for an agent loop? LongCat 2.5, on every line of the rate card — roughly 3.3 times cheaper on input, 2.25 times on output and an order of magnitude on cache reads. Whether it is cheaper per finished task is a different question and an unanswerable one without your own measurements, because the model that finishes in fewer steps wins regardless of its token price.

Would you build on LongCat 2.5 today? For an internal tool or an experiment where the price is the point and the failure mode is a retry, yes. For a customer-facing product with a service-level expectation, the combination of a provisional rate card, no deprecation window and an untested capability profile is three risks that compound rather than offset. Step 5 Preview has the opposite profile: a published licence, a downloadable checkpoint and a score of 29, which is not a high score but is a real one.

The comparison comes down to what you are buying with the difference. Three times the input price buys you a measured model with weights and a licence. If that difference is not worth it to you, LongCat 2.5 is a serious candidate and the only way to find out whether it is a good one is to run it.