Step 5 Preview vs Nex N2.5 Pro: One Number You Can Check, One You Cannot

The number circulating for Nex N2.5 Pro is 82.7, its vendor-reported score on Terminal-Bench 2.1. The number circulating for Step 5 Preview on a benchmark with the same name is 33%. Those two figures are not comparable, and the fact that they are routinely placed side by side is the clearest illustration of the problem this matchup has: one model has been measured by someone other than its maker, and the other has not been measured by anyone else at all. Step 5 Preview went live on StepFun's API on 20 September 2026 and published its weights on 15 October. Nex N2.5 Pro was announced on 8 September and published its Apache-2.0 checkpoints on Hugging Face within days of that announcement. Both are downloadable. Only one is scoreable from the outside.

That asymmetry is not a knock on Nex N2.5 Pro's engineering. It is a statement about the evidence, and evidence is what a procurement decision runs on.

What Nex N2.5 Pro is now that the weights are up

A lot of coverage still describes Nex N2.5 Pro as a model whose weights are "coming soon", because that is what its card said in the first week of September. The checkpoints have been public since mid-September: a 397-billion-parameter mixture-of-experts model with roughly 17 billion active per token, post-trained from a Qwen 3.5 lineage, under Apache-2.0. Anyone describing it as unreleased is working from a stale page.

What has not changed is the measurement situation. The vendor's benchmark figures — 82.7 on Terminal-Bench 2.1, 61.2 on SWE-bench Pro, 89.7 on BrowseComp, 87.4 on OSWorld-G — were produced through NexCUA, the alliance's own computer-use harness, which it says it intends to open-source and has not. No independent evaluator has put a page up for the model. That absence is not an accident of timing; it is a gap that has persisted through the whole period the weights have been downloadable.

The comparison that can be made

Everything below comes from published sources rather than a shared run, and each line names where the figure came from. That is the honest ceiling for this pair.

• Parameters — Step 5 Preview: 600B total, 27B active. Nex N2.5 Pro: 397B total, ~17B active.

• Context — Step 5 Preview: 1M tokens. Nex N2.5 Pro: 262,144 tokens, with a documented output ceiling near 236K.

• Modality — Step 5 Preview: text, image and video in, text out. Nex N2.5 Pro: text and image in, text out, with vision positioned as the computer-use feedback channel.

• Reasoning control — Step 5 Preview: thinking is always on, but there is a low/medium/high effort dial. Nex N2.5 Pro: none, medium (adaptive, default) and high tiers, so reasoning can be switched off entirely.

• Weights — Step 5 Preview: BF16 and FP8 under the StepFun Model License 1.0. Nex N2.5 Pro: Apache-2.0.

• Price — Step 5 Preview: $1.00 in, $2.70 out, $0.10 cached per million tokens on StepFun's rate card. Nex N2.5 Pro: a free hosted tier, with no list price published by the vendor.

The licence line is the one place where Nex N2.5 Pro is unambiguously simpler. Apache-2.0 is a known quantity to any legal team; StepFun's licence carries commercial thresholds that require modelling. The context line goes the other way, and it is close to a fourfold difference — meaningful if your agent reads whole repositories or document sets.

The comparison that cannot be made

Three things would be needed to put these two on the same axis, and none of them exists.

First, a shared benchmark version. Terminal-Bench 2.1 and Terminal-Bench 4.0 are different tests with different task sets and different scoring, and a score on one says nothing about the other. The 82.7 figure is on 2.1; the 33% figure is on 4.0. Anyone presenting them as a head-to-head is comparing a sprint time to a marathon split.

Second, a common harness. Step 5 Preview's figures come from a published index that runs a fixed evaluation suite against every model it lists. Nex N2.5 Pro's come from software its own vendor wrote and has not released. Even if the benchmark names matched, the scaffolding would not.

Third, an independent reading of any kind. There is no page for Nex N2.5 Pro on the evaluators that publish an intelligence index. We checked that directly rather than inferring it from a missing search result, and the answer today is the same as it was in September: nothing.

What can be said without overstepping is narrower and still useful: Nex N2.5 Pro claims strong agentic performance on benchmarks where Step 5 Preview's independently measured results are mid-table. That is a reason to test it, not a reason to believe it.

What a free endpoint is actually worth

Nex N2.5 Pro's hosted tier costs nothing, which converts a price comparison into a risk comparison. There is no published list price, so there is nothing to model against and no signal about what the vendor thinks the model is worth. Third-party listings carry prices in the low tens of cents per million input tokens depending on the host, and they contradict each other badly enough that none should be treated as the vendor's position.

Free also has an availability cost. Trackers that watch the free endpoint report intermittent availability — a third-party observation, not a vendor figure, and the sort of thing that is meaningless for a prototype and disqualifying for a production path. A model you cannot depend on is not cheap; it is unpriced risk with a zero on the invoice.

Against that, Step 5 Preview's $0.54 blended rate and $2.35 per completed index task are expensive in absolute terms and completely legible. You can forecast a bill from a published rate card. You cannot forecast one from a free tier whose terms have not been written down.

Step 5 Preview's record is different, not better

It is worth being precise about what Step 5 Preview does have, because "it has a score" undersells the difference and "it is better" overstates it. Step 5 Preview scores 29 on the current version of the Artificial Analysis Intelligence Index — not the 44 that launch coverage quoted, because that index version has since been replaced. It sits sixth of seventeen in its comparison class, runs at 179 output tokens per second, and costs $2.35 per index task. Those readings are same-run, reproducible and produced by an evaluator with no commercial interest in the outcome.

What that buys you is a baseline. You know where the model sits relative to fifty-odd others measured the same way, which means you can make a first-pass decision without renting a GPU. What it does not buy you is a guarantee that Step 5 Preview will beat Nex N2.5 Pro on your workload. On coding and automation, Step 5 Preview's independent results are unremarkable — 12% on FrontierCode 2.0, 33% on Terminal-Bench 4.0 — and a model with a strong unpublished agentic claim is exactly the kind of model that might beat those numbers. The point is that you would be finding out for yourself rather than reading it off a card.

Who each belongs to

Nex N2.5 Pro belongs to teams building computer-use and browser agents who are willing to run their own evaluation before committing, and who value an Apache-2.0 licence and a 17B active footprint that fits modest hardware. The free endpoint makes the trial cheap; the absent list price makes the production decision hard. Treat the vendor benchmarks as a hypothesis with a plausible mechanism behind it — a Qwen-lineage MoE post-trained on screen-reading tasks — rather than as a result.

Step 5 Preview belongs to teams whose workload is long-context and knowledge-heavy, and who need a price they can forecast. It is the more expensive and more legible option, and its licence asks a question at very large scale that Nex's does not.

Neither is in OrcaRouter's catalogue today, so this piece cannot point you at a cheaper endpoint for either. What the platform does offer around a decision like this is the rest of the stack on one key: 200-plus models at provider list price with no markup, so a vendor's rate change reaches you the same day, and automatic failover so that a model you are still trialling never becomes a single point of failure. The pattern that works for an unproven model is to route a small share of real traffic to it behind a proven fallback, measure, and only then move the share. That is cheaper than a benchmark bake-off and far more informative than a vendor card.

If you need an answer today, Step 5 Preview is the one you can justify in a design review, because every number attached to it can be checked by someone in the room. If you are building a computer-use agent and can absorb the evaluation work, Nex N2.5 Pro is the more interesting bet — cheaper to try, harder to defend, and genuinely unmeasured. The mistake to avoid is treating 82.7 as a fact and $1.00 as a disadvantage without noticing that only one of those two numbers has a second source.