On September 20, 2026, StepFun released Step 5 Preview, its new-generation flagship base model, with the news picked up by ai-bot.cn that day. It deserves a serious look: not another "bigger parameter" base model, but a sparsely activated MoE flagship built for real-world agentic tasks. It has 600B total parameters with only 27B activated per inference, a native 1M-token context, and native multimodality (text plus vision). It scored 44 on the Artificial Analysis Intelligence Index, among the top three open-weight models, and its per-task cost is claimed to be only one-eighth of Opus 5. This article takes no sides and makes five arguments: a quick look at the release and hard specs, how to read "44 points for one-eighth the money," what 22-hour continuous agent tasks actually mean, three reasons to wait for the open-source schedule and two risks, and a round of cold thinking. For the open-source coding agent companion from this batch, see our MiniMax Code CLI resource roundup; for the terminal coding agent comparison, see our terminal coding agent comparison review; for the SOP of wiring the model into a site-building flow, see our Qoder Sites SOP.
One: Release and Hard Specs, 600B Total, 27B Active, Million Context
Start with what it actually is: Step 5 Preview is StepFun's new-generation flagship base model for real-world agentic tasks, not a pure chat model and not a single-point tool. Per ai-bot.cn's September 20 collection, it uses a sparsely activated MoE architecture: 600B total parameters but only 27B activated per inference, a 1M-token context, and native multimodality (text plus vision). This differs from dense models that fire the whole stack every time: at inference it wakes only the experts relevant to the current task, separating flagship-level ability from lower per-call compute cost.
The 1M-token context is the foundation this time. In the official framing, 1M tokens can natively hold the history of execution records, tool returns, and debug feedback across a task chain of many hours, avoiding the fragmentation from chunking. For long-horizon agent tasks, chunking is the old pain point: a wrong cut loses surrounding context and every later plan sits on incomplete input. A native long context makes "run it all at once" possible, the hardware premise behind its "continuous autonomous tasks" claim.
There are three layers of access. The web tier talks directly on studio.stepfun.com for first probing the boundary. The domestic API goes through platform.stepfun.com and the overseas API through platform.stepfun.ai; you create an API key and call the interface per the docs to wire the ability into your system or agent workflow. The official site is stepfun.com/step-5-preview. The three tiers separate "model ability" from "delivery flow": the model plans and executes, external tools land the result, a division that returns below.
One note on modality scope. The native input today is text plus vision, with no audio and no video, so it can read documents, screenshots, and interfaces and iterate from visual information, but "watching video" and "listening to speech" are not native yet. Vision is a real differentiator for the Chinese market, while audio-video-native scenarios are not its main field, so do not read the scope wide. For the broader multimodal picture, compare with our Qwen3.8-Omni-Flash omni-modal hotspot.
Two: How to Read "44 Points for One-Eighth the Money": The Real Meaning of the Efficiency Route
This section is the first core argument. Per ai-bot.cn's September 20 collection, Step 5 Preview scored 44 on the Artificial Analysis Intelligence Index, among the top three open-weight models and just behind GPT-6 Astra and Opus 5, while its per-task cost is claimed to be one-eighth of Opus 5. Reading "44 points" and "one-eighth cost" together, the conclusion is not "it beat Opus 5" but "it took a Pareto route different from the closed flagship."
First, break down where the one-eighth comes from. Per ai-bot.cn's September 20 collection, the core reason is the sparse MoE: 600B total but only 27B activated per call, and "calling experts on demand" lets flagship-level ability avoid paying full price every time. This route is the opposite of "make the model bigger and more expensive"; it pursues "the same ability for less money." So the one-eighth is not a marketing discount pulled from nowhere but a structural consequence of the architecture choice, its cheapness from architecture, not subsidy.
Then compare wins and losses with Opus 5 item by item. Per ai-bot.cn's September 20 collection of the competitor sheet: on context window it is 1M tokens for Step 5 against 200K for Opus 5, a big Step 5 win and a key point for long-horizon tasks. On GPU kernel optimization it is 508 TFLOPS for Step 5 against 493 for Opus 5, a slight edge. On AIME24 post-training it moved from 53.3 percent to 60 percent, tied with Opus 5 but with fewer tokens for Step 5. On the Intelligence Index, Opus 5 scores slightly above 44, so Opus 5 wins that item. On open source, Step 5 previews full weights on October 15 while Opus 5 stays closed with API only. Put these together: Step 5 leads on cost, context, and per-unit compute, trails on absolute intelligence index, and differs in kind on "whether open" (one previews openness, the other stays permanently closed).
One caveat: these numbers come from a vendor sheet, benchmarks do not represent every real scenario, and a single score is even less equal to product experience. In particular, "one-eighth per-task cost" is a vendor-framed comparison; the real bill also depends on your task length, call frequency, and context occupancy. They are laid out here to present "near-flagship intelligence for far less money" flatly, not to cheer. For real deployment, run a small sample on your own data first. For the real cost math of long context, see our long-context agent cost review.
Three: What 22-Hour Continuous Agent Tasks Mean: The Real Boundary of Long-Horizon Autonomy
This section is the second core argument and the one needing cold water. Per ai-bot.cn's September 20 collection, Step 5 Preview can run continuous autonomous tasks for over 22 hours, such as optimizing a GPU kernel to 508 TFLOPS and beating Opus 5, while the detail page says it "can run complex tasks continuously within 24 hours." The two figures disagree, and this article states it honestly: the brief says "over 22 hours," the detail page says "within 24 hours." Whichever you take, both point to the same thing: it can keep planning, calling tools, reading feedback, and self-correcting across a task chain of many hours, instead of finishing one round and stopping.
Why does long-horizon autonomy matter? Because real engineering tasks are never single-round Q and A. Optimizing a GPU kernel, automating a post-training pipeline, or building a project from scratch all need the "plan, execute, observe, correct" loop run for dozens of rounds. In the past, models "forgot what they did earlier" on long tasks; once the context filled, they lost the thread and everything after went wrong. Step 5 keeps the whole task chain's history in the 1M-token window and, per the official disclosure, forms an autonomous "plan, execute, observe, correct" loop that reaches external tools such as serial ports, cameras, and APIs, and keeps fixing code from the real environment's returned status and errors until the task completes. This is the key difference from models that "answer only one round."
But the boundary must be clear. First, the figure itself is inconsistent; 22 hours and 24 hours are two numbers, meaning the vendor has not given a strictly audited continuous runtime, so reading it as "verified 24-hour unattended" is over-reading. Second, "continuous execution" is not "one-shot success." In a chain of many hours, any abnormal tool return, environment change, or goal drift can send later actions off course, and fixing such errors still relies on a human. Third, the real-hardware operation has a concrete case: per ai-bot.cn's September 20 collection, it can operate serial ports and cameras and keep debugging from the real hardware's returned errors for over 3 hours, with the ESP32 board retrofit as the representative. The 3 hours is an empirically anchored point, the 22 or 24 hours a higher-level claim, and the two differ in strength, so do not mix them.
State the boundary plainly: Step 5 pushes "long-horizon agent" from demo toward "usable," closer to "a strong assistant that works many hours straight" than to "a closed-door automatic factory." For practitioners, split long tasks into observable stages, keep human confirmation at key nodes, and do not solder the whole chain to one model.
Four: Three Reasons to Wait for the Open-Source Schedule and Two Risks: Weights on October 15
This section must set the fact straight first. Per ai-bot.cn's September 20 collection, Step 5 Preview previews releasing full weights on October 15, 2026, which means it is now in the Preview stage, the weights are not yet out, and it is not yet an open-source model. There is no standalone GitHub repository this time, so by site discipline this is a hotspot piece, not an open-source piece. Remember one line: what you experience on studio and the API today is the Preview, and downloading weights waits until October 15.
Three reasons it is worth waiting. First, once the weights truly ship, you can deploy locally, keep it private, modify the architecture, and distill, upgrading "use the API" to "own the model yourself," which matters most for data-sensitive and finance or government scenarios. Second, once the 600B total, 27B activated sparse MoE is open, the community can run reproducible efficiency research, and the "one-eighth cost" claim gets third-party verification rather than staying at vendor framing. Third, its design ideas on long context and the agentic loop, once open, become a reference baseline for followers and push the domestic base-model efficiency route forward.
But two risks must be stated. First, "preview" is not "certain on time." In large-model open-source history, delays, staged releases, and shipping a small model first then a big one are not rare; the October 15 date is vendor framing, and whether it lands on time, whether what ships is the full weights and under what license all need verification on the day, so do not treat it as an already-happened fact. Second, the real usability of hardware-operation tasks is pending proof. That it can debug an ESP32 for over 3 hours is empirical, but how far it generalizes to real industrial devices beyond serial ports and cameras, different wiring, and different firmware has no public benchmark. In other words, "can operate real hardware" is now a conclusion with an anchor but limited scope, not "can connect to any hardware."
One line to close: the October 15 date is worth marking, but as "to be verified," not "already confirmed." Wait until the weights truly ship, the license is clear, and the community produces reproduction results before writing it into your technical selection; that is the steady engineering posture.
Five: Cold Thinking: The Gap Between Scores and Delivery, and Two Unknowns at the Preview Stage
The final round of cold thinking, and it must splash. First, there is a gap between benchmark scores and real delivery. 44 points, 508 TFLOPS, AIME24 at 60 percent, 553 repositories across 33 languages: these numbers are bright, but all come from vendor framing or vendor-chosen scenarios. The Intelligence Index is a composite intelligence proxy, not the delivery quality of that one task in your project; AIME24 is a math contest, not production robustness; 553 repositories is vendor coverage, not a one-shot fix in your private codebase. Reading leaderboard advantage as delivery advantage is the easiest way domestic flagships mislead people. Press it with your own small sample before concluding.
Second, the Preview stage has two unknowns that cannot be answered now. One is API price. This time only the "one-eighth of Opus 5 per task" comparison framing was given, with no per-token unit price, tier price, or long-context billing method you would pay to connect. For those wiring it into production, the real monthly bill is the decision basis and that number is absent now. The other is stability. Preview means still iterating; rate limits, availability, tool-call consistency, and checkpoint recovery only show after large-scale public testing. Binding a key chain to a Preview now is at your own risk. For the landing method of wiring the model into a site-building agent workflow, see our Qoder Sites SOP; for the horizontal ability of terminal coding agents, see our terminal coding agent comparison review.
Third, competitors will not sit still. Opus 5 still leads slightly on absolute intelligence index, GPT-6 Astra sits ahead, and Alibaba, Tencent, ByteDance, and Zhipu all have long-context or agent moves. What Step 5 Preview leads in today is "using sparse MoE to drop cost to one-eighth while keeping near-flagship intelligence," and whether rivals close this window fast depends on how far they fill the efficiency route. For practitioners, wire today's ability into workflows while keeping a multi-model switch, not binding chains to one vendor. It leads on efficiency today and someone else may lead elsewhere tomorrow; keeping an exit is engineering maturity.
The article closes on one line: Step 5 Preview's meaning is not that it pushed a benchmark to first place (it did not), but that it put "near-flagship intelligence" and "one-eighth cost" on one table for the first time with sparse MoE, and showed with a 22-hour-class agent a step of the domestic base model toward "can actually work." Whether it becomes your production tool is left to the October 15 weights and your scenario; but that "the efficiency route is becoming the main battlefield of domestic flagships" is a trend StepFun spelled out for the industry on September 20.