A generated hero title card reading 'Qwen4-Max vs Step-5 Preview' with the subtitle '48% more capability, 33% more money, 6% per index point' and three pill badges reading '43 vs 29 index', '.27 vs .35 per task' and '16 November 2026'. A footer line reads 'Scores per Artificial Analysis Intelligence Index v5.0.1; prices per each vendor's rate card.' Flat vector styling on a white background with a blue-to-cyan gradient wash and the OrcaRouter logo composited bottom right.
Guides & Insights

Qwen4-Max vs Step-5 Preview: Fourteen Index Points, Half a Cent per Point

Author

Elias Hawthorne

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Divide what each model costs per task by what it scores, and the two sides of this comparison land within six per cent of each other. Qwen4-Max, which Alibaba released today, 16 November 2026, scores 43 on the Artificial Analysis Intelligence Index v5.0.1 at a measured $3.27 per index task — about 7.6 cents per index point. Step-5 Preview, StepFun's 600-billion-parameter sparse mixture-of-experts model, released on 20 September 2026, scores 29 on the same index at $2.35 per task — about 8.1 cents per point. One model is 48% smarter on the only ruler both have been held against. The other is cheaper on both sides of the meter, by a third on input and 40% on output. And per unit of measured capability the two land within seven per cent of each other, with the brand-new flagship marginally ahead.

That is not the comparison either vendor is selling, and it is the one that actually decides a migration. It also comes with an asymmetry nobody's launch post mentions: one of these two models has an independently measured output speed and the other does not have a measured speed at all yet, because it is hours old. What follows separates what is documented from what is measured, works the caching arithmetic that moves the price gap from 33% to something else entirely, and ends on the two axes where the answer is not close — weights and output ceiling.

What each one is, as of today

Qwen4-Max is Alibaba's fourth-generation flagship, built on a 2.4-trillion-parameter backbone with 71 billion parameters active per token plus 328 billion of N-gram embedding parameters. It accepts text, images and video and returns text. The documented context window is 1,000,000 tokens, with a 991,808-token input ceiling in non-thinking mode and 983,616 in thinking mode — the difference is the reasoning headroom. Maximum output is 131,072 tokens with thinking off; with thinking on, the reasoning and the answer share a 262,144-token budget. Thinking is on by default with a 131,072-token budget at a default reasoning effort of high, selectable across low, high and xhigh. Those figures are Alibaba's own documentation, and Qwen4-Max publishes no weights, does not support fine-tuning in any region, and does not support model deployment in any region either.

Step-5 Preview is StepFun's flagship for agentic work: 600 billion total parameters, 27 billion active, a 1M-token context window, and text, image and video input with text output. Reasoning effort is selectable across low, medium and high. Maximum output is documented at 64,000 tokens, with max_tokens defaulting to no cap so the model chooses the length. It supports streaming, tool calling, JSON Schema output and prompt caching, and the API is OpenAI-compatible.

Three months after launch, the "Preview" in the name has outlived its meaning. StepFun still lists it as Preview in its model overview today, while the BF16 and FP8 checkpoints have been on Hugging Face since 15 October under the StepFun Model License 1.0. Treat the label as a naming habit, not a warning that the endpoint will move under you — but do not read it as a GA commitment either, because the vendor has not made one.

The price gap is a third; the cache arithmetic is what changes it

At list prices the two are easy to summarise: Step-5 Preview is cheaper on both sides of the meter, and by more than a rounding error.

• Input — Qwen4-Max $1.50 per million tokens in Singapore, $1.413 in the other five regions, versus Step-5 Preview $1.00

• Output — Qwen4-Max $4.50, versus Step-5 Preview $2.70

• Cache read — Qwen4-Max $0.075 per million, a 95% discount, versus Step-5 Preview $0.10, a 90% discount

• Cache write — Qwen4-Max $1.875 per million for a five-minute TTL and $3.00 for one hour, versus Step-5 Preview nothing, because cache-miss input at $1.00 already includes writing the prefix

• Blended, on Artificial Analysis's 7:2:1 cache-hit/input/output mix — Qwen4-Max $0.80 per million versus Step-5 Preview $0.54

That cache-write line is the one that gets skipped and it is the one that can reverse the ranking on the right workload. Take a 300,000-token document reviewed in four passes with 3,000 output tokens each, at Singapore rates, no thinking. On Qwen4-Max the first pass writes the prefix at $0.5625 and emits $0.0135 of output; the next three read it at $0.0225 apiece plus $0.0405 of output between them. Total $0.684. On Step-5 Preview the first pass costs $0.30 of input and $0.0081 of output; the next three cost $0.03 each plus $0.0243. Total $0.422. Step-5 Preview is 38% cheaper on that job.

Collapse it to a single pass and the gap widens rather than closes — $0.576 against $0.308, with Step-5 Preview 46% cheaper — but the reason changes. On one pass there is no read to amortise, so you are paying Qwen4-Max's cache-write premium for a prefix you will never reuse. The practical rule: Qwen4-Max's caching rewards a prefix read many times and punishes one read once. Step-5 Preview's flat $0.10 cache read and write-included miss price is the friendlier structure for workloads where you cannot predict the hit rate.

A generated two-column scoreboard titled 'Qwen4-Max vs Step-5 Preview — the scoreboard'. The Qwen4-Max column reads Intelligence Index 43, cost per index task .27, cost per index point 7.6 cents, cache write .875 per 1M, weights none published, max output 131,072 tokens. The Step-5 Preview column reads Intelligence Index 29, cost per index task .35, cost per index point 8.1 cents, cache write included in .00 miss price, weights BF16 and FP8 on Hugging Face, max output 64,000 tokens. A footer reads 'Index per Artificial Analysis v5.0.1; prices and limits per Alibaba and StepFun documentation, November 2026.'

Fourteen points, and an evidence gap that runs the other way

Both models have been run through Intelligence Index v5.0.1, which incorporates nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2, and Terminal-Bench-Science 0.1. Neither model page publishes separate scores per reasoning effort, so these are single figures at whatever setting Artificial Analysis used.

• Index — Qwen4-Max 43, ranked 38th of 81 in its class, versus Step-5 Preview 29, ranked 10th of 29 among large open-weight models. The class medians differ (41 and 23), which is why those ranks are not comparable to each other even though the scores are

• Cost per index task — $3.27 versus $2.35

• Cost per index point — 7.6 cents versus 8.1 cents, so the more expensive model per task is the cheaper one per unit of capability

• Output speed — Step-5 Preview 124 tokens per second, versus no published figure for Qwen4-Max. Artificial Analysis lists its time per task as unknown and its provider page reports no provider benchmarks at all

• Time to first answer token — Step-5 Preview 23.53 seconds, versus nothing published for Qwen4-Max

• Time per index task — Step-5 Preview 15.2 minutes versus the 13.3-minute median for its class; Qwen4-Max has no figure

• Verbosity — Qwen4-Max emitted 127 million output tokens across the index against a 78.6-million median, versus Step-5 Preview's 129 million against a 104-million median. Both are verbose; Qwen4-Max is verbose relative to a leaner peer group

The 23.53-second first token is the number that should worry anyone planning an interactive deployment of Step-5 Preview. It is not a networking artefact: it is what a 600-billion-parameter model with thinking on by default costs you before the answer starts. Qwen4-Max's reasoning defaults — thinking on, 131,072-token budget, effort set to high — point the same direction, and until somebody measures it there is no honest way to say which is slower. That is the practical cost of buying on launch day: you get a score and a rate card, and you do not get a latency number.

Alibaba's own evaluation material is a separate matter and should be read as vendor-reported, unreproduced. On Work360, its in-house benchmark spanning 26 top-level industry domains, the vendor reports 52.69 for Qwen4-Max against 47.74 for Qwen3.8-Max. It also reports average cost per Work360 task falling from $4.43 to $1.67, a 62% reduction, and token use falling about 24% at xhigh effort and 35% at high effort on that benchmark, with 48% and 49% reductions on Terminal-Bench 4.0. Those are claims about Qwen4-Max relative to its predecessor. They say nothing about Step-5 Preview, and none of them has been independently checked.

A screenshot of the Artificial Analysis model page for Qwen4 Max, captured 16 November 2026, showing 'Released November 2026', the Alibaba Cloud provider label, an Intelligence Index of 43 ranked 38 of 81, in-benchmark ranks for performance, cost and verbosity, .50 input and .50 output per 1M tokens, a 95% cache discount, .27 cost per task, 127M tokens generated, and a 984k-token context window.

Weights: this one is not close

Step-5 Preview is downloadable today. The Hugging Face repository carries BF16 and FP8 safetensors under the StepFun Model License 1.0 — an Apache-2.0-style licence with an attribution requirement and a commercial-use condition — plus the chat template, three quantised variants and deployment recipes: FP8 on four B300s or eight H200s under vLLM, FP8 on the same hardware under SGLang, BF16 on eight B300s or sixteen H200s. About 1,600 downloads in the last month.

Qwen4-Max has no checkpoint and no path to one. Alibaba's documentation marks fine-tuning as unsupported in every region and model deployment as unsupported in every region, which is a stronger statement than "the weights are not out yet" — it says the vendor is not going to let you run this model yourself through its platform at all. If your requirement is a model you can quantise, fine-tune, air-gap or pin to a specific build, the comparison ends here and it ends in StepFun's favour.

It also changes what "open" buys you. Step-5 Preview's licence permits commercial use with conditions, so a legal review is part of the adoption cost, and the deployment recipes assume a four-to-sixteen-GPU node — this is not a model you run on a workstation. But it is a model whose behaviour cannot be changed under you by a silent provider update, and for anyone building a regulated or long-lived system that is worth more than fourteen index points.

The output ceiling, and a documentation conflict worth knowing

This is the least glamorous difference and the one most likely to bite a specific workload. Qwen4-Max gives you 131,072 tokens of output with thinking off and a shared 262,144-token budget with thinking on. Step-5 Preview's API documentation says 64,000 tokens of maximum output. Its Hugging Face model card says output can run to the full 1M tokens, with input and output counting against the same limit.

Those two statements come from the same vendor within a month of each other, and neither is marked deprecated. If you are planning a long-output task — a full report, a large patch, a generated document — read the API documentation rather than the card, and test the ceiling before you architect around it. The 64,000 figure is consistent with the 15.2-minute time per index task; the 1M figure would imply something quite different.

One more asymmetry worth flagging: both models take video input, but only StepFun publishes the operational detail. Up to 60 images per request at low or high detail, video as URL, Base64 or a stepfile:// reference, MP4 under 128 MB with clips under five minutes recommended. Alibaba documents video input for Qwen4-Max without publishing comparable limits on the model page. If video is central to your workload rather than incidental, that difference in documentation quality is itself a signal about which vendor has been asked the question before.

Who should pick which, this week

Pick Qwen4-Max if the ceiling matters more than the bill: hard reasoning, long-horizon agentic work, or anything where 43 against 29 on a nine-evaluation index is the difference between a task that completes and one that does not. Accept two risks in exchange. The first is that you are buying on day one with no independent latency figure and no third-party reproduction of the vendor's Work360 claims. The second is the cache-write premium, which quietly taxes single-pass long-context work at $1.875 per million tokens written.

Pick Step-5 Preview if you need to own the model, if you need a measured 124 tokens per second and a 23.53-second first token you can plan around, or if your workload is cache-heavy in a way you cannot predict. Accept that you are buying 29 index points, that the first token arrives slowly, and that the output ceiling is documented two different ways by the vendor.

The third option is the honest one for most teams this week: neither, exclusively. The gap between these two models per unit of measured capability is 6%, which is well inside the error bar of a single index version, and both have been evaluated only once by one harness. A team that commits to one of them on the strength of a launch-day score is taking a position on a number that has not been reproduced.

Running both without a second contract

Neither model is in the OrcaRouter catalogue today — 215 model identifiers, checked this morning, with no Qwen4 entry and no StepFun text model among them. Both are reachable through their own vendors' APIs: Alibaba's six regions for Qwen4-Max, StepFun's OpenAI-compatible endpoint at api.stepfun.ai/v1 for Step-5 Preview. Saying otherwise would be the kind of claim that ages badly within a week.

What we do route is the model Qwen4-Max replaces. Qwen3.8-Max is in our catalogue as qwen/qwen3.8-max at Alibaba's own list price of $2.00 input and $6.00 output per million tokens, passed through with no markup, alongside 214 other identifiers on one OpenAI-compatible base URL. That matters for exactly one reason in this comparison: a launch-day model is the worst possible thing to bet a production path on, and the sane posture is to keep the proven generation on the route while you measure the new one against your own traffic. The routing DSL expresses that as a policy rather than a code change — send the hard, reasoning-heavy calls to the new model, keep the high-volume work on the cheaper proven one, and let automatic failover catch the case where a brand-new endpoint degrades under load you did not predict.

Pass-through pricing also means the arithmetic in this article stays honest. When Alibaba cuts a rate — as it did today, moving Qwen4-Max to $1.50 and $4.50 from Qwen3.8-Max's $2.00 and $6.00 — a catalogue that passes the provider's list price through reflects it the same day rather than at the next contract renewal. That is not an argument for us over either vendor's own API. It is an argument for not signing anything that locks a price you have not seen hold for a month.

A screenshot of the OrcaRouter model page for qwen/qwen3.8-max, captured 16 November 2026, showing the model identifier, text, image and video input with text output, a 1M-token context window, a list date of 2026-08-03, .00 input and .00 output per 1M tokens, measured latency figures, and an OpenAI-compatible Python sample pointing at https://api.orcarouter.ai/v1.

The comparison readers will want in a fortnight is not the one in this headline. It is Qwen4-Max's latency and Step-5 Preview's stable release, both measured by someone who is not the vendor, against workloads people actually run. Until then the defensible position is narrow and unglamorous: a 48% capability gap, a 33% price gap, a 6% gap per unit of capability, and a set of decisions — weights, output ceiling, cache structure — that no index score captures and that will outlast this launch week.