StepFun's 600B Step 5 Preview Ties Kimi K3 on Artificial Analysis at 64% Lower Cost

TL;DR

StepFun launched Step 5 Preview on September 20: a 600B-parameter mixture-of-experts model with 27B active per token, a 1M-token context window, and text, image and video input. On the Artificial Analysis Intelligence Index it scores 44, the same score as Kimi K3 (max), at $0.72 per index task instead of $2.00. The API is live at $1.00 in and $2.70 out per million tokens, and StepFun promises open weights on October 15. The weights did not wait. Within two hours of the announcement, an outside account copied StepFun's own Hugging Face repo, all 1.2 TB of it, and ten public copies were online by Monday.

What StepFun shipped

The launch post on X calls Step 5 Preview StepFun's "new flagship model for agentic work," with "particular strength in finance." The developer docs fill in the practical limits:

• Size: 600B total parameters, 27B active per token.
• Context: 1M tokens in, 64K tokens out.
• Input: text, up to 60 images per request, and video (MP4, QuickTime, Matroska, with clips under five minutes recommended).
• Controls: reasoning effort at low, medium or high, tool calling, JSON Schema output, and prompt caching.
• APIs: Chat Completions plus a Messages API, and StepFun documents a Claude Code setup for its Step Plan subscription.

The official rate card lists $1.00 per million input tokens on a cache miss, $0.05 on a cache hit, and $2.70 per million output tokens, with reasoning tokens billed as output. StepFun's previous open release, Step 3.7 Flash, lists at $0.20, $0.04 and $1.15. So the flagship costs five times more on input and still undercuts most of the field.

The number that matters: cost per task

Per-token prices are a sticker. What you actually pay is price multiplied by how much the model talks, and reasoning models talk a lot. Artificial Analysis measures that directly with cost per index task: the weighted average cost to finish one task in its suite, counting input, cache, reasoning and answer tokens.

Think of per-token price as the price per liter at the pump and cost per task as what the road trip actually cost. Cheap fuel is no bargain if the engine drinks twice as much of it.

Step 5 Preview needed 160M output tokens to run the index, well above the 92M median for its peer group, so it is not a terse model. Kimi K3 (max) used the same 160M. That makes the gap against Kimi purely a pricing story: Kimi lists at $3 in and $15 out, and a full index run cost $3,658.07 against $924.68 for Step 5.

GLM-5.3 (max) scores one point higher at 45, but it burned 210M output tokens and $2,503.48 getting there. StepFun's own launch chart draws exactly this comparison against GLM-5.3 and Kimi K3 and labels it "-65% Cost." The independent numbers put it at 64%, so the vendor chart holds up.

Where it sits on the curve

Step 5 Preview does not top the leaderboard and does not claim to. GPT-6 Astra (max) leads the comparison at 53 for $3.26 per task, Claude Opus 5 (max) scores 51 at $5.86, and GPT-5.6 Sol (max) gets 47 for $1.99.

The pitch is the middle of the curve. Step 5 lands about 93% of GPT-5.6 Sol's score for 36% of its cost per task. The only model in this comparison that is cheaper per task is DeepSeek V4 Pro 0813 (max) at $0.67, and it scores 36. If your agents grind through thousands of mid-difficulty tasks rather than one hard proof, that slot is the one that shows up on the invoice.

The weights showed up 25 days early

StepFun's announcement ends with a date: "Open weights on Oct 15." Hugging Face's own metadata tells a faster story.

According to a snapshot by OrcaRouter, StepFun's official repo, stepfun-ai/Step-5-Preview-BF16, was created at 03:52 UTC, 37 minutes after the X post. At 05:07 UTC, an account that does not appear in StepFun's public Hugging Face member list created a repo whose first commit reads "Duplicate from stepfun-ai/Step-5-Preview-BF16." That commit carries a co-author credit for a member of the stepfun-ai organization.

The copy holds 28 safetensors files totaling 1.215 TB, which is what 600B parameters weigh at two bytes each in BF16. By that afternoon, OrcaRouter found the official repo reduced to a single .gitattributes file. As of Monday morning UTC, Hugging Face shows no public repo at that address at all. A Chinese developer forum thread titled "Step 5 Preview weights leaked, now withdrawn" (our translation) picked up the copy the same day, and one reply offered the whole diagnosis: when you publish, you are supposed to pick Private.

The copies kept multiplying anyway. We counted ten public repos named Step-5-Preview-BF16, each exactly 1.215 TB, with the latest created at 01:35 UTC on September 21. We found no public statement from StepFun about any of them.

The practical advice does not change. The direct copy contains no license file and no model card, and StepFun has not said what license Step 5 will ship under. Its last two open releases, Step 3.5 Flash and Step 3.7 Flash, both use Apache 2.0, while the two models Step 5 ties on the index, Kimi K3 and GLM-5.3, are listed by Artificial Analysis as commercial-use restricted. Do not build a product on unlicensed weights 25 days before the licensed ones are due.

What the config file gives away

The copied config.json is the first public look at the architecture, and it lines up with StepFun's headline numbers. Our back-of-envelope count from the config lands at about 598B total parameters and 26.5B active for the text stack, excluding the vision encoder and prediction heads.

• 92 layers, 4,096 wide. 88 of them are mixture-of-experts layers.
• 352 routed experts, 8 per token, plus a shared expert. This is how 600B stored becomes 27B computed.
• Mostly local attention. Three of every four layers use a 512-token sliding window. The other 23 attend globally through a sparse indexer that keeps the top 512 candidates instead of scoring every token.
• Three multi-token prediction layers, the kind of extra heads used to speed up decoding with speculative drafts.
• A 47-layer vision encoder, shipped as its own file.

The sparsity is easiest to picture as a hospital with 352 specialists. Every patient gets triaged to eight of them plus the on-call generalist, so the building is enormous but each visit is cheap. The attention layout applies the same thrift to the 1M-token context: most layers only look back 512 tokens, and the global layers work from a shortlist.

One more detail: the model is marketed as Step 5, but inside the config the text model's class is named Step4ForCausalLM. Version numbers are marketing. Class names are archaeology.

Can you run it?

Not on a gaming rig. The BF16 checkpoint is 1.2 TB before any KV cache. By simple arithmetic, an 8-bit build would be around 600 GB and a 4-bit quant around 300 GB, which puts it in multi-GPU server or very large unified-memory territory, and a long context adds cache on top. The 27B active parameters help speed once the model is loaded, not the memory needed to load it.

Until October 15 the realistic path is the API. StepFun supports both OpenAI-style Chat Completions and a Messages API it documents for use with Claude Code, so pointing an existing harness at it is mostly a base-URL change. The $0.05 cache-hit rate is 95% off input, which matters most for agents that resend the same repo or document set on every turn.

Caveats

• One scoreboard. The Intelligence Index is independent of StepFun, but a tie at 44 is a tie on one aggregate, not proof of equal performance on your workload.
• It is a preview. Behavior and pricing can change before or after the open release.
• The architecture read comes from a copy. Hugging Face metadata ties it to StepFun's repo and the numbers match StepFun's stated specs, but StepFun has not published its own model card yet.
• No vendor benchmarks here. Every score in this post is from Artificial Analysis, not from StepFun's self-reported tables.

Key Takeaways

• Step 5 Preview is StepFun's new flagship: a 600B mixture-of-experts model with 27B active, 1M context, and text, image and video input, live on the API at $1.00 in and $2.70 out per million tokens ($0.05 cached).
• It ties Kimi K3 (max) at 44 on the Artificial Analysis Intelligence Index for $0.72 per task against $2.00, a 64% cut. GLM-5.3 (max) scores 45 at $2.01.
• The gap is price, not brevity. Step 5 and Kimi K3 both used 160M output tokens to run the index.
• StepFun's own BF16 repo was copied within two hours of launch. Ten 1.215 TB copies were public by Monday, and the official repo no longer shows publicly.
• Wait for October 15. The copies carry no license, StepFun's earlier open models used Apache 2.0, and the Step 5 license is still unannounced.

Sources: StepFun launch post on X, StepFun Step 5 Preview page, StepFun developer docs: Step 5 Preview, StepFun pricing and rate limits, Artificial Analysis: Step 5 Preview, Artificial Analysis: Kimi K3 (max), Artificial Analysis: GLM-5.3 (max), Artificial Analysis: GPT-5.6 Sol (max), OrcaRouter, locdd forum thread, stepfun-ai on Hugging Face