Step 5 Preview vs GLM-5.2: A Four-Month Track Record Against a Two-Week-Old Endpoint

Ask a working engineer to choose between Step 5 Preview and GLM-5.2 and most will reach for the newer one, because the newer one is cheaper per token and scores ten points higher on the independent board. Ask the same engineer to put it in front of paying users and the answer gets harder, because one of these models has been carrying production traffic since June 16, 2026 and the other one was released on September 20, 2026. Step 5 Preview is the faster, cheaper, multimodal, higher-scoring model. GLM-5.2 is the one whose failure modes are already written down somewhere by somebody else, whose weights ship under MIT rather than a custom grant with revenue thresholds in it, and whose rate card has not moved in four months. Neither of those is a tiebreaker you can put in a spreadsheet, and together they are the whole decision. What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The spec comparison is genuinely lopsided and it is worth getting out of the way first, because the argument for GLM-5.2 is not a spec argument. Z.ai's model is text-only, 753 billion parameters total with roughly 40 billion active per token, capped at 128,000 output tokens, and priced at $1.40 per million input and $4.40 per million output. StepFun's is $1.00 and $2.70, takes video as well as text and images, caps a response at 64k, and runs 27 billion active parameters. On the numbers a buyer sees first, StepFun wins five rows out of six.

Every row of the contrast

• Input price — Step 5 Preview $1.00 per million tokens vs GLM-5.2 $1.40, a 29% edge to StepFun

• Output price — $2.70 vs $4.40, a 39% edge, and the meter that dominates an agent's bill

• Cached input — Step 5 Preview $0.10 per million, a 95% discount vs GLM-5.2 $0.26, an 81% discount

• Cost per completed Intelligence Index task — Step 5 Preview about $0.71 vs GLM-5.2 $2.01, which folds verbosity and cache behaviour into one number and is the least flattering reading of StepFun's card anywhere in this comparison

• Independent score — Step 5 Preview 44 vs GLM-5.2 34 on Artificial Analysis Intelligence Index v4.3.2, where GLM-5.2's page also carries a deprecation marker

• Context and output — 1M tokens of context for both; 64k maximum output for Step 5 Preview against 128K for GLM-5.2

• Modality — Step 5 Preview text, images and video in, text out vs GLM-5.2 text in, text out, with no image path at all

• Architecture — Step 5 Preview sparse mixture-of-experts, 600B total and 27B active vs GLM-5.2 753B total with roughly 40B active, an active-parameter figure Z.ai does not state on the model card and which third-party listings supply

• Licence — StepFun Model License 1.0, a custom grant with an attribution trigger and separate-commercial-licence thresholds vs GLM-5.2 under MIT, with no threshold, no attribution requirement and no negotiation path to model

• Weights — both downloadable; Step 5 Preview in FP8 and BF16 since release day plus an NVFP4 build published October 14, 2026 vs GLM-5.2 in BF16 and FP8 through the zai-org organisation

• Rate-card stability — Step 5 Preview unchanged since September 20 vs GLM-5.2 unchanged since June 16

Two notes before those rows get over-read. The 44-versus-34 gap is not the same comparison it looks like. GLM-5.2 launched in June with a headline index figure of 51, which was produced under an earlier revision of the same index; Artificial Analysis has since restated the scale, and the 34 above is the current-revision reading of a model the evaluator now marks deprecated. Quoting 51 against 44 would be comparing two different rulers, and it is the single most common error in coverage of this matchup. Second, StepFun's model has one scored configuration — a single 44 at one effort setting — while GLM-5.2 has been measured across a longer window, so the newer model's number is a point where the older one has something closer to a profile.

The licence difference is the one that survives contact with a legal review

Both of these models are described as open-weight, and the two documents are not equivalent. GLM-5.2 is MIT. That is a licence a lawyer reads once and approves permanently: use, modify, distribute, sublicense, sell, no attribution obligation, no user threshold, no separate agreement for any business model this side of absurdity. Weights are published through Z.ai's own Hugging Face organisation in BF16 and FP8, the model card states no regional limits, and there is no clause anywhere in it that a startup has to model before it can plan a product.

Step 5 Preview's terms are more permissive than no weights and more restrictive than MIT, and the gap is not rhetorical. StepFun Model License 1.0 grants a perpetual, worldwide, royalty-free, non-exclusive right to use, copy, distribute, sublicense, modify and create derivative works, with a matching patent grant. Two conditions attach. A commercial product or service incorporating the model that passes 100 million monthly active users or $20 million in revenue in any calendar month must display "Powered by StepFun AI" prominently in the product interface and in the user agreement. A hardware manufacturer or brand owner shipping the model's capabilities in devices, or a model-as-a-service business whose combined revenue exceeds $20 million over any twelve months, needs a separate commercial licence. Internal use that does not expose the model or its capabilities to a third party sits inside the grant without any of that.

For a team building an internal tool, this difference costs nothing and the argument for GLM-5.2 on licence grounds is weak. For a team building a hosted product with a revenue curve they expect to cross $20 million, it is the difference between a legal checkbox and a line item with a counterparty attached, and it is the kind of thing that gets discovered after the architecture is committed. The honest framing is that StepFun's licence is a standard shape — Llama and other open-weight licences use similar thresholds — and that MIT is simply a shorter conversation.

What a track record is actually worth

GLM-5.2 has been in the world since mid-June. In that time it has run inside agent harnesses, been reproduced by third-party hosts and reviewers, accumulated publicly reported failure modes, and had two successors shipped on top of it — GLM-5.3 on August 18, 2026, on the same base weights, and GLM-5.5 after that. Z.ai never yanked the pricing and never withdrew the weights. That last point is more important than it sounds: the practical risk of adopting an older model is not that it is worse, it is that it becomes unsupported while you are still running it, and GLM-5.2 has now survived two generational refreshes with its rate card intact.

Step 5 Preview's operational record is measured in weeks, and the specific thing it lacks is the unglamorous one: a body of reports about how it behaves under sustained load, at high concurrency, with pathological prompts, and across a long enough window for a vendor-side regression to have happened and been noticed. Neither StepFun's model card nor its documentation is a substitute for that, and the vendor's own benchmark table — presented as an image rather than as quotable text — is not an independent measurement of anything. The two harnesses do have numbers attached to them: StepFun names the settings it evaluated at, including a temperature of 1.0, top-p of 0.95, top-k of 0 and the SWE-agent harness for DeepSWE v1.1. That is more disclosure than many release cards offer. It is still the vendor grading its own work.

There is one narrow exception where the newer model's numbers are genuinely independent and worth trusting: the 44 on the Intelligence Index, which is the only neutral score Step 5 Preview has, alongside the roughly $0.71 per completed index task and the evaluator's "very verbose" label on its 160-million-token suite run. Treat the 44 as a single well-sourced number and not as a profile. Treat the verbosity label as the more likely thing to bite you, because it is a statement about cost rather than quality and it shows up on every invoice.

Where the choice actually lands

Take GLM-5.2 if any of these is true. You need an MIT licence with no thresholds, because your legal team will otherwise spend a quarter on it. You need a 128,000-token single response, which Z.ai's model provides and StepFun's does not at any price. You are running a text-and-code workload where a model with four months of public production history behind it is worth more than ten index points, because you would rather inherit somebody else's bug list than generate your own. Or you need the cheapest possible self-hosted option and a 753B Mixture-of-Experts with a published IndexShare sparse-attention design that Z.ai says cuts per-token FLOPs by roughly 2.9 times at full context, and you intend to actually deploy it rather than call it.

Take Step 5 Preview if the workload wants video, or if the ten-point index gap is the difference between a demo that works and one that does not, or if you simply want the cheapest token on the page and are willing to be an early adopter. The video capability is real but unscored — StepFun documents MP4, QuickTime and Matroska input by URL, Base64 or a Files API reference, with URL-based files under 128MB and a recommended length under five minutes, and no independent evaluation of it exists. That makes it a reason to run a test rather than a reason to commit, and the same is true of the whole model: its strongest rows are the ones its own vendor is the only source for, and its most trustworthy row — the index — is the one where it loses to nothing and wins by ten.

The two proposals are not equally risky in the other direction either. Choosing the older model costs you capability you can measure. Choosing the newer one costs you an unknown you cannot, and the way to bound an unknown is to not bet the production path on it. That is the case for running Step 5 Preview behind a fallback rather than instead of something: route a slice of traffic at it, keep GLM-5.2 or whatever else you trust as the fallback, and let cost per completed task decide rather than the sticker. On OrcaRouter, GLM-5.2 sits at z-ai/glm-5.2 at Z.ai's list price with 0% markup — $1.40 and $4.40 with the $0.26 cache read passed through unchanged — and because the pass-through is the provider's own price, a Z.ai rate change reaches your bill the day it lands rather than at the next cycle. Automatic failover is what makes a two-week trial of an unfamiliar endpoint safe: the trial model fails, the request does not.

Step 5 Preview is not in OrcaRouter's live catalogue, and the honest instruction is to call it through StepFun's own platform or self-host it from the published FP8, BF16 or NVFP4 checkpoints. For this particular comparison that is not much of a loss, because the reason to pick it over GLM-5.2 is that you want to be early, and being early usually means living with one provider anyway.

The one thing that would settle this comparison is not a benchmark. It is time. In three months Step 5 Preview will have the thing GLM-5.2 has now: a public record of how it behaves under other people's load, at other people's concurrency, in other people's failure cases. Until then the choice between these two models is less about which is better and more about which kind of evidence you are willing to act on — the evidence you inherited, or the evidence you generate yourself.