Step 5 Preview vs MiniMax M3: Cheaper Per Token, Pricier Per Point

The unit price favours MiniMax M3 by a wide margin. At $0.30 per million input and $1.20 per million output, the vendor's 428B sparse model undercuts Step 5 Preview's $1.00 and $2.70 by more than a factor of two on both sides. But unit price is the wrong denominator for a model you use to answer questions, and once you divide by what each model actually delivers, the ranking flips. Step 5 Preview costs $0.71 per Intelligence Index task at a score of 44. MiniMax M3 costs $0.51 per index task at a score of 29. Per point of measured capability, the model with the higher sticker price is the cheaper one — and the reason is a fifteen-point index gap that no amount of discounting closes.

Two very different kinds of "available"

MiniMax M3 has been generally available since June 1, 2026, when MiniMax opened it through MiniMax Code, its Token Plan and the API — no preview label anywhere in the launch. The weights followed on June 12, eleven days later, as promised. The repository MiniMaxAI/MiniMax-M3 carries roughly 428 billion total parameters with about 23 billion active per token, built on MiniMax Sparse Attention, alongside a separate MXFP8 quantised release.

Step 5 Preview opened its API on September 20, 2026. Its weights have not been published — the Hugging Face repository stepfun-ai/Step-5-Preview-BF16 is a bare shell — and StepFun has committed to a date of October 15, 2026. So the practical difference is that M3 is a model you can download, self-host and inspect, while Step 5 Preview is a hosted endpoint with a published intention.

Worth flagging: MiniMax's own model page still says M3 "will soon be fully open-sourced," which contradicts the repository it published in June. Treat that as a stale vendor page rather than evidence the weights are missing.

The price story is not the cost story

Artificial Analysis computes a cost-per-index-task figure by running the same evaluation set through each model and totalling what the tokens cost. It is the closest thing the field has to a total cost of ownership number for a fixed amount of work, and it folds in the two effects that token prices hide: how verbose a model is, and how much thinking it does.

Step 5 Preview lands at $0.71 per index task, with a total index run cost of $922.84. MiniMax M3 lands at $0.51, with a total run cost of $537.98. On raw throughput of tasks, M3 is about 28% cheaper. Divide by index score — 44 against 29 — and Step 5 Preview comes out at roughly 1.6 cents per index point against M3's 1.8 cents.

A caveat belongs right here: that is one index run per model, and an index run is a small sample. The per-point arithmetic is a sanity check on the per-task figure, not a benchmark in its own right. But the direction is consistent with the per-task numbers, and those are the figures the two vendors' own price lists cannot show.

• Intelligence Index — Step 5 Preview 44 (#24 of 200) vs MiniMax M3 29 (#15 of 113), both on the current index version

• Input price — $1.00 per million vs $0.30 per million

• Output price — $2.70 per million vs $1.20 per million

• Cost per index task — $0.71 vs $0.51

• Active parameters — 27B vs 23B

• Context window — 1M tokens vs 1M tokens

• Output speed — 99.8 tokens/s vs 210.4 tokens/s

• Weights — promised October 15, 2026 vs available since June 12, 2026

The index-version trap that makes this confusing

There is a genuine hazard in reading M3's numbers, and it is worth a paragraph because it trips up almost every third-party comparison page. Artificial Analysis recalibrated its Intelligence Index on September 7, 2026, and the recalibration compressed scores across the board. Under the current scale M3 measures 29. Under the pre-September scale it measured 44 to 45 — and a number of trackers still publish 44.4 or 45.4 for it.

That collision matters more than usual here, because Step 5 Preview's current score is also 44. A page that puts a stale M3 figure beside a current Step 5 Preview figure appears to show two models tied, when the current gap is fifteen points. Both figures in this article are from the current index version and are comparable to each other; anything published before September 7 is not.

Where MiniMax M3 is genuinely hard to beat

Two things about M3 are not close. The first is speed: 210.4 output tokens per second, fourth fastest of the 113 models Artificial Analysis tracks, against a median of 76.3. Step 5 Preview's 99.8 tokens per second is respectable and less than half of it. M3's time to first token of 1.01 seconds is likewise the faster figure against 2.96 seconds.

The second is input modality. M3 accepts text, image and video and returns text, and MiniMax documents desktop and computer-use operation. Step 5 Preview takes text and images. If your pipeline feeds screen recordings or video frames into a model, that is not a preference — it is a capability difference.

The price gap reinforces both. At $0.30 and $1.20 with a permanent 50% discount applied to a list of $0.60 and $2.40, and an 80% cache discount that brings cache reads to $0.06 per million, M3 is one of the cheapest ways to move a large volume of tokens through a capable model. Note the tiering: input above 512K tokens doubles the whole request, and the priority service tier costs 1.5×.

Where it comes apart

MiniMax's own evaluation set for M3 is strong and should be read as vendor-reported: SWE-bench Verified at 80.5%, SWE-bench Pro at 59.0%, Terminal-Bench 2.1 at 66.0%, MCP Atlas at 74.2%, BrowseComp at 83.5 and OSWorld-Verified at 70.06%. MiniMax ran these on its own infrastructure using Claude Code as the scaffolding, averaged over four runs, and the scoring setup is not public — so they are not reproducible by a third party as things stand.

The independently measured picture is much harsher on agentic work. Artificial Analysis recorded Terminal-Bench 4.0 at 2% for M3 — a different version of the benchmark from the vendor's Terminal-Bench 2.1 figure, and not comparable to it, but stark regardless. SciCode came in at 47%, AutomationBench-AA at 21%, and CritPt at 4%. M3's strongest independent results are on long-context and knowledge work: AA-LCR v1.1 at 83%, Humanity's Last Exam at 39%.

Step 5 Preview inverts that profile. Terminal-Bench 4.0 at 33.3% on the same harness as M3's 2%, SciCode at 59% against 47%. If the work is agentic or code-shaped, this is not a close comparison and the price advantage does not compensate for it.

The licence clause most comparisons skip

MiniMax M3 is open-weight, but it is not open source in the OSI sense, and the terms are stricter than the phrase "open weights" suggests. The repository carries the MiniMax Community License: free for non-commercial use, with commercial use conditioned on prominently displaying "Built with MiniMax M3" — and any product with more than $20 million USD in annual revenue requires prior written authorisation from MiniMax. It is not MIT, not Apache, and not the kind of licence a legal team waves through.

Step 5 Preview cannot be assessed on this axis at all yet, because there is no licence to read. StepFun has committed to a weights release on October 15 and has not published the terms that will govern it. Anyone planning to build on it should treat the licence as an open question rather than assume it will resemble Moonshot's or DeepSeek's.

This is the honest asymmetry between the two: M3's terms are restrictive but known; Step 5 Preview's are unknown. A restrictive licence you can plan around beats an unannounced one you cannot.

Calling either of them

MiniMax M3 is in our catalogue at /models/minimax/minimax-m3, reachable through the same key and request shape as the rest of the 200-plus models we route, with provider list price passed through at 0% markup — so the permanent 50% discount MiniMax has applied is the price you see, and it moves the day MiniMax moves it. Automatic failover is the other half of that: M3 at 210 tokens per second is attractive for high-volume work, and high-volume work is exactly where a single degraded endpoint hurts most.

Step 5 Preview is not in our catalogue. StepFun's text models are not among those we serve, so it is available through StepFun's own API and several third-party platforms — and it is not reachable on the key you would use for M3.

The verdict

Take MiniMax M3 when volume, latency or video input is the constraint, when you want a model you can download today, and when you can live inside the community licence — including the $20 million revenue threshold. Accept that independent agentic scores put it a long way behind, and that its vendor benchmarks have not been reproduced outside MiniMax's own infrastructure.

Take Step 5 Preview when the work is agentic or code-heavy, when you want the capability per dollar rather than the token per dollar, and when you are comfortable being an early customer of a hosted endpoint whose weights arrive next month. Accept that you cannot check the licence before you build, and that a 2% versus 33.3% Terminal-Bench 4.0 spread is the kind of gap that decides a project.

The discount is real. It just is not the thing being bought.

MiniMax M3 is one of the models we route at 0% markup with automatic failover, so a vendor price change is live on the same key the same day.