Step 5 Preview vs Kimi K3.1 Pro: Both Models Were Re-Scored Last Week. Here Is What Survives

If you looked up Step 5 Preview and Kimi K3.1 Pro a week ago you would have come away with two numbers: 44 and 55. Both are gone. On 27 and 28 October 2026, Artificial Analysis replaced its Intelligence Index with version 5, which takes 60% of its score from private or held-out test sets, and the same model scored under the old rubric and the new one is not the same measurement. Step 5 Preview reads 29 under v5. Kimi K3.1 Pro reads 46. The nine-point gap between them is now a comparison you can make honestly, and it is a different gap from the one the launch coverage was written around.

What follows is that comparison, plus the part nobody puts in a comparison table: the two models are on opposite sides of a weights question that resolves on 4 November, and only one of them can be called through OrcaRouter today.

The two numbers everybody quotes are both wrong now

Kimi K3.1 Pro launched on 23 October 2026 with an Intelligence Index score of 55 and a fourth-place debut, eleven points above Kimi K3. Four days later it read 46. Nothing about the model changed; the ruler did. Artificial Analysis rebuilt the composite into nine evaluations across four categories — Knowledge Work, Coding, Science and General — and dropped three of the older ones, including GDPval-AA v2.1, SciCode and AA-LCR v1.1. It added FrontierCode 2.0, a set of 132 private coding tasks graded on whether a patch is worth merging, and Terminal-Bench-Science.

Step 5 Preview took the same hit, harder. It reads 29 under v5.0, against 44 under the previous index, and it loses more of its supporting evidence than Kimi does, because StepFun's launch material leaned on exactly the evaluations that were removed. The 88.3 on AA-LCR and the 58.9 on SciCode that appear in most coverage of this model are now scores from benchmarks that no longer contribute to the headline index. They are still true measurements of something. They are no longer the thing the index measures.

The ranking consequence is worth stating plainly: under v5, an open-weights-or-promised-open-weights model at 46 is 12 points behind the leader at 58, where under v4 the gap had been 3. That is a statement about the rubric, not about a sudden collapse in capability, and any article that presents it as the latter is misreading its own source.

The spec lines, one dimension at a time

• Total and active parameters — Step 5 Preview: 600B total, 27B active per token, 8-of-352 experts plus one shared. Kimi K3.1 Pro: 2.8T total, 104B active.

• What kind of release it was — Step 5 Preview: a new flagship shipped in preview form, with weights following. Kimi K3.1 Pro: a post-trained version of Kimi K3 with larger post-training compute, including joint reinforcement learning across hundreds of domains — not a new architecture.

• Context window — Step 5 Preview: 1M tokens. Kimi K3.1 Pro: 1,048,576 tokens.

• Input modalities — Step 5 Preview: text, images and video. Kimi K3.1 Pro: text and images.

• Input price per million tokens — Step 5 Preview: $1.00. Kimi K3.1 Pro: $2.00, flat across the full context window with no long-context surcharge.

• Output price per million tokens — Step 5 Preview: $2.70. Kimi K3.1 Pro: $10.00.

• Cached input — Step 5 Preview: a 90% discount on the input rate. Kimi K3.1 Pro: $0.20 per million to read, with cache writes at $2.00 on a five-minute TTL and $4.00 on a one-hour TTL.

• Weights — Step 5 Preview: published 15 October 2026, BF16 and FP8, under StepFun Model License 1.0. Kimi K3.1 Pro: not published; Moonshot has said weights are expected by 4 November 2026.

• Independent score under Intelligence Index v5.0 — Step 5 Preview: 29. Kimi K3.1 Pro: 46.

Cost per finished task, with the caveat attached

Artificial Analysis reports $2.35 per Intelligence Index task for Step 5 Preview and $1.98 for Kimi K3.1 Pro. That looks like a small edge for Kimi, and it is worth understanding why it is not obviously one.

The two figures were not produced on the same task set, because the task set was rebuilt between the two measurements. Both are "cost per index task" under whichever composite was live when the number was taken, which makes the comparison directional at best. What is not ambiguous is the mechanism: Kimi K3.1 Pro is verbose, spending around 55,000 output tokens per task and 150 million across a full run against a median of 87 million, and at $10 per million output that verbosity is expensive. Its $1.98 is a consequence of finishing in fewer billable steps, not of a cheaper token.

Step 5 Preview is the cheaper token by a factor of 3.7 on output and 2 on input, and the faster model by a wide margin on the measurement available: 179.1 output tokens per second on StepFun's API against a class median of 74.1, where Kimi K3.1 Pro measures around 73. If your workload is latency-sensitive and output-heavy, that ordering matches the price ordering and the decision is easy. If your workload is one where the answer quality difference of 17 index points matters more than the time to first token, the same numbers point the other way.

Weights: one is out, one is a promise with a date

This is the asymmetry that the benchmark tables hide. Step 5 Preview's weights have been downloadable since 15 October: BF16 and FP8 checkpoints on Hugging Face and ModelScope, vLLM and SGLang recipes on the same day, a technical blog describing the 92-layer narrow-deep stack and sparse GQA attention pattern. The FP8 checkpoint wants roughly 773 GB of VRAM and at least four GPUs, with B300 as the default target and eight H200s on the Hopper path. The licence is StepFun's own, not OSI-approved: attribution on materially used outputs, a separate negotiation for hardware makers, and a commercial licence for model-as-a-service operators above $20 million in annual revenue.

Kimi K3.1 Pro's weights did not exist on launch day — Artificial Analysis still classified the model as proprietary on 24 October — and the announced date is 4 November 2026. When they land, Moonshot's model becomes the highest-scoring downloadable model on the v5 index, which is a genuinely significant event for anyone who wants frontier-adjacent capability without a vendor in the loop.

Until 4 November, the comparison is between a model you can run and a model you can only call. That is not a permanent disadvantage for Kimi, but it is a real one this week, and it is the reason a decision made today may not be the decision you would make in nine days.

Where each one is callable right now

One of these two models is on OrcaRouter and the other is not. Kimi K3.1 Pro is in the live catalogue as kimi/kimi-k3.1-pro, alongside five other Kimi models. Step 5 Preview is not: the catalogue carries 214 model IDs across 16 providers, and there is no StepFun entry in it as of 29 October 2026. If you want Step 5 Preview today, you call StepFun's own API or you host the checkpoint.

That asymmetry is the useful part of this matchup for a routing decision. Because we pass through provider list price with no markup, the Kimi rate card you see here is Moonshot's rate card — $2.00 and $10.00 — and the same is true of any cut Moonshot makes later. And because failover is automatic, the practical version of "Kimi is the cheap, measured option" is that you can put it first in a chain and let the platform fall back when a provider degrades, rather than encoding a fallback into your client and maintaining it.

The migration question, if you are switching

Three things will cost you time regardless of which direction you move.

The first is output volume. Kimi K3.1 Pro's cost profile is dominated by how much it writes, not by its rate. A prompt that produces a tight answer on Step 5 Preview may produce three times the tokens on Kimi, and your bill will follow the tokens rather than the rate card. Measure output tokens per task on your own workload before you compare the two prices at all.

The second is cache economics. Step 5 Preview discounts cached input by 90%. Kimi K3.1 Pro charges $0.20 per million to read a cache but also charges to write one — $2.00 at a five-minute TTL and $4.00 at an hour. An agent loop that re-sends a long system prompt every step is a very different proposition on those two schemes, and the cheaper-looking card is not always the cheaper loop.

The third is harness sensitivity. Moonshot's own table puts Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0 against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra; the independent evaluator reports 67% for the same model. Terminal-Bench numbers move by tens of points with scaffolding, so a 1.7-point gap between two labs' runs is not a measurement of a 1.7-point capability difference. Treat every cross-vendor row in this comparison as a claim about direction.

What to do this week, concretely: if your workload is long-context, output-heavy and latency-sensitive, Step 5 Preview at $1.00 and $2.70 is the cheaper experiment and the weights make it a portable one. If the work is agentic coding where patch quality is the whole product, Kimi K3.1 Pro's 46 on v5 and its 67% on Terminal-Bench 4.0 make it the stronger buy, and 4 November is the date that turns it from a call into an asset. Either way, do not build a decision on the 44 and the 55. Those numbers belong to a rubric that no longer exists.