Step 5 Preview vs MiMo V2.6 Pro: One Index Point for Nine Times the Bill

On the same Artificial Analysis run, StepFun's Step 5 Preview and Xiaomi's MiMo V2.6 Pro finish one intelligence point apart: 29 against 30. They are not one bill apart. The same run prices a completed index task at $2.35 for Step 5 Preview and $0.26 for MiMo V2.6 Pro, a gap of roughly nine to one. Both are open-weight mixture-of-experts models with a million-token window, both take images as well as text, and both arrived within a day of each other in late September 2026 — Step 5 Preview on StepFun's API on 20 September, MiMo V2.6 Pro on 21 September. That makes this the cleanest head-to-head available in the open-weight tier, and it is the one comparison where the cheaper model is also the marginally better-scoring one.

What follows uses a single evaluation run, taken on the index version Artificial Analysis is serving today, because almost every comparison page still on the web uses a superseded one. The numbers below are reproducible, same-run, and label who produced them.

Two release notes, one week apart, two very different weight drops

The timelines diverge as much as the architectures. Step 5 Preview reached an API on 20 September and then spent nearly four weeks as a hosted-only product before StepFun published the checkpoints on 15 October. MiMo V2.6 Pro shipped its MIT-licensed weights alongside the API on 21 September, and Xiaomi published its reinforcement-learning recipe, roughly 7,000 task environments and a distilled 9B sibling at the same time.

The spec sheets, one line per dimension:

• Total and active parameters — Step 5 Preview: 600B total, 27B active per token. MiMo V2.6 Pro: 1.0T total, 42B active.

• Architecture — Step 5 Preview: 92 layers, sparse grouped-query attention with 8-token block indexing, and a 3:1 mix of gated sliding-window attention with sparse full attention. MiMo V2.6 Pro: sparse MoE with 384 routed experts, eight active per token.

• Context and output — Step 5 Preview: 1M tokens. MiMo V2.6 Pro: 1M tokens, with a 131K output ceiling on our own model page.

• Input — Step 5 Preview: text, image and video. MiMo V2.6 Pro: text, image, video and audio, and it advertises native audio rather than video-only.

• Weights and licence — Step 5 Preview: BF16 and FP8, under the StepFun Model License 1.0. MiMo V2.6 Pro: MIT.

• Price per million tokens — Step 5 Preview: $1.00 in, $2.70 out, $0.10 cached. MiMo V2.6 Pro: $0.435 in, $0.87 out, $0.0036 cached.

The parameter gap is the first thing to notice. Step 5 Preview is the smaller model by total size and the smaller one by active compute, 27B against 42B. That is not an accident of packaging; StepFun's technical blog argues the opposite trade — a narrow, deep stack trained to behave like a much larger model on long-horizon agent work, which is why 92 layers sit on a 600B envelope instead of a trillion-parameter one.

The board was re-scored, and the old numbers are still everywhere

If you search for this matchup today, most results will tell you MiMo V2.6 Pro scores 46 and Step 5 Preview scores 44. Those were real readings. They were taken on an earlier version of the Artificial Analysis Intelligence Index, and that version has since been replaced. On the board as it stands now, the same two models read 30 and 29, and the models that used to sit near them have moved by similar amounts.

This matters for a practical reason rather than a pedantic one: every derived figure computed from the old scores is now wrong. The "$0.13 versus $0.72 per task" ratio that circulated at launch has become $0.26 against $2.35. The ordering did not change — MiMo still leads by one — but the magnitude of everything around it did, and the direction of the cost story changed from a rounding error to a decision.

A caution on how to read any of these numbers: the index is a weighted composite, and its sub-benchmarks disagree sharply for this pair. Reporting the composite alone would hide the only part of the comparison that gives Step 5 Preview a reason to exist.

Where the nine-to-one buys something

Three sub-results from the same run are genuinely in Step 5 Preview's favour, and they cluster around a single theme: saying less when it does not know.

• Humanity's Last Exam (Diamond) — Step 5 Preview 24% vs MiMo V2.6 Pro 15%. Nearly a nine-point lead on the hardest knowledge benchmark in the set.

• AA-Omniscience — Step 5 Preview 16 vs MiMo V2.6 Pro −4. The score rewards calibrated abstention as well as correct recall, and a negative value means a model that answers confidently and wrongly more often than it stays quiet.

• Output speed — Step 5 Preview 179.1 tokens per second vs MiMo V2.6 Pro 58.0, measured on the same service. Three times the throughput is a real latency difference in an agent loop that makes a dozen sequential calls.

Those three together are a coherent defence of the premium: a model that hallucinates less and returns faster is worth more per token in exactly the workloads both vendors are targeting — long-running agents, repository-scale prompts, research synthesis. Whether it is worth nine times more is a separate question, and the answer depends on how much a wrong confident answer costs you.

Where MiMo wins without an asterisk

The code and automation tests go the other way, and not narrowly.

• FrontierCode 2.0 — MiMo V2.6 Pro 24% vs Step 5 Preview 12%. Double the score on the coding benchmark, which is the single largest divergence in the pair.

• AutomationBench-AA — 59% vs 51%.

• Terminal-Bench 4.0 — 35% vs 33%, a narrow lead that matters because Step 5 Preview's agentic positioning is built on terminal work.

• GDPpdf — 19% vs 15%, and Terminal-Bench-Science 6% vs 2%, both modest but consistent.

Step 5 Preview's wins on HLE and Omniscience read as breadth of knowledge and calibration; MiMo's wins on FrontierCode and AutomationBench read as the ability to act. For a coding agent, those are not equivalent currencies.

The cache economics are the part most likely to surprise someone modelling spend. MiMo V2.6 Pro's cached input rate is $0.0036 per million tokens against Step 5 Preview's $0.10 — a 28-fold difference on the token class that dominates a long agent session, where the same repository or document context is re-read on every turn. MiMo also sells an UltraSpeed tier at $4.35 in and $8.70 out for latency-critical calls, which is a way to buy throughput without changing models.

What an agent turn actually costs

Take one representative call: a 400,000-token context in, 20,000 tokens out. On list prices that is $0.454 for Step 5 Preview and $0.191 for MiMo V2.6 Pro — a 2.4-fold gap, not a nine-fold one. Now run the same context ten times in a session. With the cache doing its job, Step 5 Preview re-reads those 400,000 tokens at $0.10 per million and MiMo V2.6 Pro at $0.0036, and the gap on the cached portion widens to roughly 28 to one.

That arithmetic is why the per-task gap and the sticker gap disagree. The per-task figure is a weighted average across a full evaluation with heavy prompt reuse, so it reflects the cache rate as much as the headline rate. If your workload is one-shot prompts with fresh context every call, budget nearer the 2.4-fold figure. If it is a persistent agent re-reading the same files, the nine-fold figure is the more honest planning number.

What you can call today

MiMo V2.6 Pro is in OrcaRouter's catalogue under the model id xiaomi/mimo-v2.6-pro, billed at Xiaomi's own list rate of $0.43 in and $0.87 out per million tokens with no markup added on top. That matters here for a specific reason: the pricing in this comparison is the thing most likely to move, and a pass-through router means a Xiaomi price change lands on your bill the same day rather than at the next contract renewal.

Step 5 Preview is not in our catalogue. Its place to call remains StepFun's own API, or a third-party host that carries the checkpoints. For a team weighing the two, that asymmetry is itself an argument: one side of this comparison can be tested against a fallback in a single routing config, and the other cannot. The routing DSL lets you send the coding traffic to MiMo V2.6 Pro and keep a second model behind automatic failover, so a spike in latency at one provider does not take the agent down. If MiMo's one-point lead turns out not to hold on your own workload, switching costs a model id, not a contract.

Questions this matchup keeps raising

Why do most comparison pages still show 46 versus 44?

Because they were written against the previous version of the index and have not been recomputed. The underlying evaluation changed, not just the models, so every figure derived from it — per-task cost included — moved with it. When two pages quote this pair and disagree by more than a point or two, check which index version produced each.

Does Step 5 Preview's calibration advantage survive a domain shift?

Partly. The Omniscience and HLE results are measured on general knowledge; the calibration behaviour they reward is most valuable where the model is asked about things outside its training distribution. In a narrow domain where both models were trained hard — code, for instance — MiMo's FrontierCode lead suggests the abstention advantage is not doing the work you are paying for.

Is either model a drop-in replacement for the other?

Not without checking. They share an OpenAI-compatible surface in practice, but the reasoning controls differ, the output ceilings differ, and Step 5 Preview's 27B active footprint changes how much of a long prompt you can keep resident. Port the prompt, re-run your own evaluation set, and compare on your workload rather than on either vendor's slide.

The short version of the decision: if your workload is knowledge-heavy, latency-sensitive, and expensive to get wrong, Step 5 Preview's nine-fold premium is defensible and its 179 tokens per second makes it pleasant to use. If your workload is code, automation and long sessions over the same context, MiMo V2.6 Pro is the better model at a price that is not close — and it is the one of the two you can put behind a router today and switch away from tomorrow.