Step 5 Preview Is Announced: What StepFun's 600B Flagship Actually Ships, and What's Still Missing

StepFun announced Step 5 Preview on September 20, 2026 — a 600-billion-parameter sparse mixture-of-experts flagship with 27B parameters active per token, a 1M-token context window, native image input, and an Artificial Analysis Intelligence Index score of 44. The API opened the same day. What did not open is the model itself: the Hugging Face repository stepfun-ai/Step-5-Preview-BF16 existed by the afternoon of the announcement and contains exactly one file, .gitattributes, with no weights, no license, no model card and no configuration. StepFun's own materials put the open-weights release on October 15, 2026. So the honest one-line summary of this launch is that the model is callable today and downloadable in about three and a half weeks, and those are two different things. This is also the same model this blog covered earlier today as an unannounced leak, benchmarked under a test tag before StepFun had published a product page — a useful reminder that a preview can move from leak to launch without a single number changing.

The repo is a placeholder, and that is the story

When a vendor publishes weights, the repository is the evidence. It has a license file, a config with a parameter count in it, a README with the intended use and the evaluation table, and safetensors shards whose total size tells you whether the parameter claim is real. stepfun-ai/Step-5-Preview-BF16 has none of that. The Hugging Face API record for it shows a creation timestamp of 2026-09-20T03:52Z, zero downloads, two likes, an empty config object, no pipeline_tag, no license and a single sibling file. The StepFun organisation page lists it, and lists no other Step 5 repository — no FP8 build, no NVFP4 build, no GGUF, none of the quantisation variants that accompanied Step-3.7-Flash in June.

Read that as scheduling, not as a mystery. The BF16 suffix in the repo name is the giveaway: a lab that intends to publish a bfloat16 checkpoint first — the format you fine-tune and quantise from, before the FP8 and NVFP4 serving builds arrive — names the repository that way at creation time and fills it later. StepFun has said October 15 in its own announcement materials. Until then the repository is a claim about an intention, and the only thing it proves is that StepFun has reserved the namespace.

This matters for a practical reason. The Preview suffix and the missing weights are the same signal pointing the same direction: the model is callable, the pricing is set, and the artifact you would run on your own hardware does not exist yet. Any plan that assumes a self-hosted Step 5 Preview before mid-October is a plan with no artifact behind it.

What was actually announced

StepFun's framing is narrow and specific: Step 5 Preview is a base model for real-world agentic work — AI coding, software engineering, professional knowledge work and finance — with the emphasis on long context, multi-turn tool calls and sustained execution rather than on chat quality. The company skipped the Step 4.x line entirely, going from Step-3.7-Flash straight to Step 5, which is itself a statement about how much of a step it considers this to be.

One dating detail is worth flagging, because it is the kind of thing that trips up a procurement calendar: Artificial Analysis dates this model to September 18, 2026, two days before StepFun's own announcement. The board scores a model when it can reach it, and that two-day gap is the ordinary lag between a model becoming callable and a vendor writing about it. StepFun's announcement — September 20, with the API and Studio opening the same day — is the date to cite, and the October 15 weights date comes from the same materials.

The specification as published:

• Total parameters — 600B, sparse mixture-of-experts

• Active per token — 27B, roughly 4.5% of the total, an unusually sparse ratio

• Context window — 1M tokens

• Input — text and images natively; output is text only

• Reasoning — yes, with extended thinking

• Price — $1.00 per million input tokens, $2.70 per million output tokens, with a 95% cache discount

• Availability — API and Studio open from September 20; weights scheduled for October 15

The efficiency claim StepFun is leading with is that the 600B/27B split puts the model on the Pareto frontier — capability per unit of compute — and the company frames it as three generations of the same pursuit, from Step-3.5-Flash through Step-3.7-Flash to this. It is a coherent story, and the sparse ratio is the mechanism: at 27B active, per-token serving cost sits in a much smaller model's band while the 600B total is mostly a memory-footprint question.

Vendor claims, and the third-party numbers next to them

Two classes of number appear in the launch materials, and they should not be read the same way. StepFun's own evaluations are the first class: a single-task cost claim of one-eighth of Claude Opus 5; second place behind either GPT-6 Astra or Claude Opus 5 on ALE-CLI, FrontierFinance and DRACO; a 24-hour GPU kernel optimisation run that lifted MLA kernel peak performance to 508 TFLOPS against Claude Opus 5's 493; an automated post-training experiment that moved Qwen3-30B-A3B from 53.3% to 60% on AIME24; DeepSWE v1.1 at 67.7% and its in-house StepCodeBench at 49.0%. StepFun built StepCodeBench itself — 553 code repositories, nine task types, 33 programming languages — which makes it a reasonable instrument for measuring its own model and not an independent one.

The second class is measured by someone else, and it is the part to plan against. Artificial Analysis scores Step 5 Preview at 44 on Intelligence Index v4.3.2, ranking it 24th of 200 models — level with Kimi K3, one point behind GLM-5.3, and level with Claude Opus 5's medium reasoning-effort setting. On Terminal-Bench 4.0 the same board records 33.3%, ahead of DeepSeek V4.1 Flash at 26.8% and well ahead of Kimi K3 at roughly 12.6%. Output speed measures 99.8 tokens per second and time to first token 2.96 seconds — both from the same independent run, and both the kind of figure that decides whether an agent loop feels interactive.

One third-party number cuts the other way and deserves to sit beside the price. The same index run used 160M output tokens for Step 5 Preview against a median of 92M across the models scored — the model is verbose. At $2.70 per million output tokens, that verbosity is not free: 160M tokens of output is roughly $432 of the $922.84 the run cost in total, and on a model charging three times as much it would be the dominant line on the invoice. A low headline price and a high token count per task are two halves of one number, and the cost per index task — $0.71 — is the one that already contains both.

What the leak got right, and the one thing it did not settle

This blog's earlier coverage was written from an evaluation run that surfaced before any announcement, and it flagged the claims it could not confirm. The announcement resolved most of them. The 600B/27B pair is now vendor-stated rather than inferred. The 1M-token context window is now in StepFun's own specification rather than only on a third-party board. The price of $1.00 and $2.70 is the price. The name is Step 5 Preview, not a rumoured alternative.

What the announcement did not do is resolve the context-window contradiction that the leak coverage raised. A separate third-party configuration circulating in developer tooling listed the model at 350,000 tokens of context with a 64,000-token maximum output. A 1M window and a 350k window are different products with different memory costs, and a 64k output ceiling is a hard constraint on exactly the long-horizon agentic work this model is being sold for. StepFun's announcement says 1M and does not mention an output cap. It is worth knowing which one your routing ends up on before you build a pipeline that depends on the answer, and that is a question for the vendor's documentation rather than for a leaderboard.

The other thing the launch did not change is independent reproduction. Every benchmark row above that is not from Artificial Analysis is StepFun's own, run on StepFun's harnesses, and no third party has reproduced the agentic results. The pattern from this year's Chinese-lab flagships is that the aggregate index holds up and the vendor's headline rows drift; treat the aggregate as the planning number.

What you can call today, and what to route in the meantime

Step 5 Preview is not in our catalogue, and OrcaRouter does not route it — there is a public API and a Studio on StepFun's side, and that is where it runs. What we do have is the set of models it is being measured against, behind one key: DeepSeek V4.1 Flash at $0.15 in and $0.60 out per million, GLM-5.3 at $1.26 and $3.96 against the $1.40 and $4.40 Z.ai lists, Claude Opus 5 at $5.00 and $25.00, and Kimi K3 alongside them. That is the practical shape of the next three weeks: a preview you can benchmark against, and a set of production models you can actually depend on, reachable through one API for more than 200 models with provider list price passed through at 0% markup — so if StepFun's pricing moves before October, the number you would pay moves with it the same day, not at renewal.

Automatic failover is worth more than usual in this window, because the failure mode of a preview is not that it is bad — it is that it is capacity-constrained. A model that opened its API on announcement day and has no weights yet is a model whose serving fleet is still being built, and a preview endpoint that degrades at 2am takes your agent loop with it. Put the preview behind a router with a proven model as the fallback and you get a quality signal on your own workload without betting a production path on an unproven fleet.

The date that matters is October 15

Everything above is provisional in one specific way: the weights are the thing that makes a model permanent. A closed preview at $1.00 and $2.70 is a good price from a single vendor, and a BF16 checkpoint under a published license is a price no single vendor controls any more. StepFun has committed to the second for October 15, and the repository name it has already reserved says bfloat16 first.

What to watch between now and then is narrow and checkable. Does the repository fill — and with what license? Does a config file confirm 600B total and 27B active? Does StepFun publish the output ceiling alongside the context window, resolving the 350k/64k question? Does the price hold once the preview loses its suffix? Those four answers decide whether Step 5 Preview becomes a model you self-host, a model you rent, or a model you benchmark once and move past. Until the repository has more than a .gitattributes in it, the announcement is the beginning of the story rather than the end of it.