StepFun released its new flagship foundation model, Step 5 Preview, on September 20. It is available immediately in the company's own products and via API, with full weights set to be open-sourced on October 15.
The model uses a sparse mixture-of-experts architecture with 600B total parameters and 27B activated per token, a 1-million-token context window, and native text and vision input. The launch materials name specific target uses: AI coding, software engineering, professional knowledge work, and finance.
Benchmark Scores and In-House Tests
On the Artificial Analysis Intelligence Index, Step 5 Preview scored 44, placing third among open-weight models worldwide. Its other public benchmark scores are: GPQA Diamond 93.5%, Terminal-Bench v2.1 85.0%, BrowseComp 88.7%, and the multimodal MMMU-Pro at 76.0%. StepFun also said the model achieved the best or second-best results on benchmarks such as AA-LCR and CyberGym.
A separate set of numbers needs to be viewed differently. StepFun also published scores on StepCodeBench and FinStepBench, two benchmarks it built in-house. StepCodeBench covers 553 code repositories, 9 task categories, 20 application domains, and 33 programming languages, testing real development tasks such as bug fixes, feature development, code refactoring, and environment configuration. StepFun itself notes in its materials that these in-house benchmarks were designed with different goals than public leaderboards, and results across different setups are not directly comparable.
On agentic capability, StepFun says the model can autonomously carry out continuous tasks lasting more than three hours, supporting code debugging, serial port access, screenshot capture, camera access, and simulated mouse operations. These claims currently lack third-party reproduction; the only figures outsiders can verify are the public benchmark scores.
A Cost Claim, and Its Fine Print
The most-quoted line from the launch is that per-task cost is just one-eighth that of Anthropic's Claude Opus 5. That figure comes from StepFun itself, and the public materials don't disclose how it was calculated — how many tasks, what difficulty distribution, whether it's based on listed API prices or actual usage. None of that is specified.
Running that ratio through a concrete scenario: if a team spends 100,000 yuan a month on Opus 5 inference, an eighth of that would put Step 5 Preview at roughly 12,500 yuan, saving a seven-figure sum over a year. That math only holds if the task mix matches what StepFun used in its own calculation — and that's exactly the part that hasn't been disclosed. For teams considering a migration, running the same tasks themselves after the weights are released on October 15 would be more reliable than quoting this multiple.
The choice of comparison also signals positioning. StepFun didn't benchmark against another open-weight model — it picked one of the highest-priced tiers among top-tier closed models.
A One-Month Window
The preview goes live via API first, with the full weights following a month later — a cadence more Chinese labs have adopted this year: claim the leaderboard and the news cycle first, release weights later. The risk is that during that one-month gap, a comparable open-weight flagship could land at any time, and no one can guarantee that the No. 44 ranking will still hold by October 15.
From an engineering standpoint, the 600B total / 27B active configuration is relatively deployment-friendly: a 27B activation footprint means single-node, multi-GPU inference is feasible, without needing to provision resources as if for a 600B dense model. That's also a precondition for open-sourcing the weights at all — push the parameter count much higher, and the practical value of open weights gets eaten up by deployment barriers.
StepFun has not published the license terms or API pricing this time. Those two factors will determine how many people actually deploy it after October 15.
Sources: StepFun official announcement, Ifeng Technology, CocoLoop, Sina Technology. Parameter counts, the Artificial Analysis Intelligence Index score, and the open-source release date have been checked against official statements; the cost multiple and in-house benchmark results are StepFun's own figures.