Step 5 Preview vs GLM-5.3: One Point Apart, and Only One of Them Has a Licence File

On the independent board, Step 5 Preview and GLM-5.3 are one point apart: 44 against 45 on Artificial Analysis Intelligence Index v4.3.2, where GLM-5.3 is scored as GLM-5.3 (max) and Step 5 Preview carries no effort label at all. That is the tightest gap in this series, and it is also the least useful number in the comparison. The number that decides things is that GLM-5.3 has weights you can download today — published by Z.ai under a custom licence that is not MIT — while Step 5 Preview, announced by StepFun on September 20, 2026, has a Hugging Face repository containing a single .gitattributes and a stated release date of October 15 for the BF16 checkpoint. StepFun's launch material describes the model as a top-three open-source result. It is not open source yet. GLM-5.3 is, and its licence is the thing to read before you plan around it.

What one index point actually buys

Both models were run through the same ten evaluations by the same lab, which makes the comparison unusually clean and unusually unhelpful at the same time.

• Intelligence Index — Step 5 Preview 44 vs GLM-5.3 45

• Where that sits — GLM-5.3 is one point above Step 5 Preview; Step 5 Preview is level with Kimi K3

• Parameters — Step 5 Preview 600B total / 27B active vs GLM-5.3 753B total / 40B active

• Output speed — Step 5 Preview 99.8 tokens/sec vs GLM-5.3 72.1 tokens/sec

• Time to first token — Step 5 Preview 2.96s vs GLM-5.3 2.99s

• Price per 1M tokens — Step 5 Preview $1.00 in / $2.70 out vs GLM-5.3 $1.40 in / $4.40 out

• Cache discount — Step 5 Preview 95% vs GLM-5.3 81%

• Cost per Intelligence Index task — Step 5 Preview $0.71 vs GLM-5.3 $2.01

• Context — both 1M tokens; our catalogue lists GLM-5.3 with a 131.1K output ceiling, StepFun has not published one

• Weights — Step 5 Preview closed, BF16 checkpoint scheduled for October 15; GLM-5.3 published under a custom Z.ai licence

Read those rows together and the one-point gap stops being the headline. Step 5 Preview is 29% cheaper on input, 39% cheaper on output, generates tokens 38% faster, and carries a cache discount 14 points deeper — and the cost-per-index-task figure, which folds verbosity and cache behaviour into a single number, is 2.8 times lower. It does that at 600B total against GLM-5.3's 753B, with 27B active against 40B. On the efficiency axis this is not a close matchup; on capability it is as close as the board gets.

The one row where Step 5 Preview does not win is the one that is hardest to fix: GLM-5.3 has a licence file and Step 5 Preview does not.

The licence is the whole difference, and it is not MIT

It is easy to assume a Chinese lab's open flagship ships under MIT, because several do. Z.ai's own GLM-5.2 and GLM-5.3-Flash both carry MIT. The GLM-5.3 flagship does not — it ships under a custom "glm-5.3" licence that permits commercial use subject to a set of restrictions. That is a materially different document from the MIT terms its smaller sibling carries, and for a company whose lawyers read the licence before the README, it is a conversation rather than a formality.

Step 5 Preview has no licence at all yet, which is a different problem and not obviously a smaller one. A restrictive licence tells you what you may do and leaves you to decide whether the terms are acceptable. No licence tells you nothing: no commercial-use grant, no redistribution right, no fine-tuning permission, and no way to evaluate whether the October 15 checkpoint will be usable for your product until it appears. The repository name StepFun reserved — stepfun-ai/Step-5-Preview-BF16 — names a bfloat16 format, which is the artefact you fine-tune and quantise from rather than the one you serve. That tells you the shape of the release. It does not tell you its terms.

So the honest comparison is between a licence you can read and disagree with, and a licence that does not exist. For a team that needs to fine-tune, self-host in a controlled environment, or ship a derivative product, GLM-5.3 is the only one of these two with an answer, and the answer is a legal review rather than a green light.

The efficiency gap is the part that holds up

Efficiency claims deserve more scepticism than benchmark claims, because they are easier to make and harder to check. These two hold up better than most, and the mechanism is visible in the architecture.

Sparse mixture-of-experts models only spend compute on the experts a token routes to, so the activated figure governs production behaviour while the total is mostly a memory question. At 27B active against GLM-5.3's 40B, Step 5 Preview is doing roughly a third less per-token work, and the measurements follow: 99.8 output tokens per second against 72.1, and a cost per index task of $0.71 against $2.01. That is the same story told three ways, which is what makes it credible rather than a single flattering number.

The cache discount is the row that matters most for the workload both models are sold for. An agent loop re-sending the same system prompt and repository context on every turn lives almost entirely inside that discount, so 95% against 81% compounds across a long session in a way the list price does not. At a 7:2:1 cache-hit / input / output mix, Step 5 Preview's blended rate lands near $0.51 per million tokens against GLM-5.3's materially higher blend — and the gap widens the longer the loop runs.

What is not yet testable is whether that efficiency survives at scale. Step 5 Preview's API opened on launch day, on a single endpoint. GLM-5.3 has been carrying production traffic since mid-August across many independent callers, and a model with five weeks of real load behind it has failure modes that are known to a large enough population that they surface publicly. A days-old endpoint has none of that. The efficiency numbers are the more favourable reading; the operational track record is the more useful one.

What you can call on each side today

This is the one part of the comparison where the two are not symmetric, and it is worth being exact about it rather than rounding to "both are available."

GLM-5.3 is live on OrcaRouter under the z-ai/glm-5.3 model page at $1.26 input and $3.96 output, against the $1.40 and $4.40 Z.ai lists on the model page itself — passed through with zero markup, so the number you see is the provider's current rate rather than one we chose. The open weights are on Hugging Face, so the same model is reachable three ways: our endpoint, Z.ai's own API, or your own hardware.

Step 5 Preview is not in our catalogue, and we do not route it. It runs on StepFun's own API and Studio, and that is the only place it runs until the weights land. What sits beside it on our side is the set of models it is being measured against, behind one key for more than 200 models: GLM-5.3 at the discounted rate above, DeepSeek V4 Pro at $0.66 and $1.98, and Claude Opus 5 at $5.00 and $25.00. That is what makes the next four weeks workable rather than blocked — you can benchmark the preview against a production model on the same key, with automatic failover across providers holding the production path while the preview's capacity curve is still unknown, and the routing DSL composing the production leg into a single call rather than a second integration.

Choosing between a lead and a track record

If you are fine-tuning, self-hosting or building a derivative product, GLM-5.3 is the answer and it is not close — not because it scores one point higher, but because it is the only one of the two with a licence to read and a checkpoint to download. Budget the legal review, because the custom terms are not MIT and the difference is the kind that shows up in a procurement audit rather than in a benchmark.

If you are calling a model by API and your workload is agentic text at volume, Step 5 Preview is ahead on everything the invoice and the clock care about: cheaper per token, cheaper per task, faster to generate, and a deeper cache discount for the repeated-prefix pattern that fills an agent loop. Against a days-old single-source endpoint with no published output ceiling and no licence, the case for making it your default is a case about a benchmark, not about a contract.

What would change the answer is the October 15 checkpoint. A permissive licence would make the efficiency lead something you can own rather than rent, and at that point the one-point deficit stops mattering — a model that is 2.8 times cheaper per index task and 38% faster at generation does not need to win the aggregate. A restrictive licence leaves GLM-5.3 as the only one of the pair you can build a product on, and the one-point gap becomes trivia about two models you would use differently anyway.

The one specification worth watching is the output ceiling. Both vendors advertise a 1M-token context. Our catalogue lists GLM-5.3 at a 131.1K maximum output; StepFun's announcement names a 1M context and no output cap, and a separate third-party configuration that circulated during the leak listed 350,000 context with a 64,000 output maximum. For a pair of models sold on long-horizon agentic work, that number decides what you can actually build, and right now only one side has answered it.