StepFun Unveils Step 5 Preview: A Mixture-of-Experts Model with 600B Parameters (27B Active) and 1M Context Window Built for Extended Agentic Tasks
StepFun launched Step 5 Preview, a flagship sparse Mixture-of-Experts model with 600B total parameters (27B active per token) and a 1M-token context window, targeting agentic workloads like coding, research, and financial analysis. It uses a deep-and-narrow 92-layer architecture, on-policy long-horizon RL training, and offers API pricing well below rivals, though its reasoning chains are notably verbose.
If you build AI agents for coding, research, or financial analysis, StepFun's latest release deserves a look. The company has launched Step 5 Preview, a new flagship model aimed squarely at agentic workloads, and its main selling point is simple: similar intelligence to top rivals at a much lower cost per task. This article breaks down the architecture, benchmark results, pricing, and what to expect before you decide whether to try it.
What Exactly Was Released?
Step 5 Preview is a sparse Mixture-of-Experts (MoE) model with roughly 600 billion total parameters, of which only about 27 billion activate per token — around 4.5% of the full network. This design keeps inference cheap while preserving capacity.
The official documentation lists these specifications:
- Model ID: step-5-preview
- Context window: 1M tokens
- Input: text, images, and video
- Output: text
- Reasoning effort levels: low, medium, and high
- Features: streaming, tool calling, JSON Mode, JSON Schema support, and prompt caching
Right now the model is available through StepFun's hosted API and platform. Self-hosting isn't possible yet — StepFun says open weights will arrive on October 15, 2026. If you want to run it yourself later, do the math: 600B parameters need roughly 1.2 TB of memory in BF16, before accounting for the KV cache, so multi-GPU server hardware will be a requirement.
On the agent side, StepFun says the model coordinated 950 web fetches in a single agent action, and the company documents a Claude Code integration through its Step Plan.
Architecture: Deep and Narrow, Not Wide
Rather than making the network wider, StepFun stacked 92 Transformer layers in a narrow-and-deep configuration (as reported by Pandaily). The team's argument is that deeper stacks create longer computational paths, which support implicit multi-hop reasoning — exactly the kind of thinking needed when an agent searches the web, runs code, and processes tool outputs during long prefill stages.
Training relied heavily on on-policy, long-horizon reinforcement learning. StepFun highlights bit-wise alignment between training and inference across MoE routing, along with several efficiency techniques:
- MTP-3 speculative decoding
- FP8 MoE computation
- KV-cache offload
Together, StepFun reports more than 3x end-to-end speedup for long-horizon RL training.
Benchmarks: Company Numbers vs. Independent Testing
One important detail up front: StepFun ran Step 5 Preview at High effort, while competitors ran at Max. Here are the company-reported results (via RuntimeWire):
| Benchmark | Step 5 Preview | Claude Opus 5 | GPT-6 Astra |
|---|---|---|---|
| FrontierFinance | 66.4 | 69.7 | 55 |
| DRACO | 83.3 | 87.6 | 76.8 |
| DeepSWE v1.1 | 67.7 | — | — |
| StepCodeBench | 49.0 | — | — |
| ProgramBench | 80.5 | — | — |
On coding, GPT-6 Astra and Claude Opus 5 stay ahead on all three coding benchmarks. Note that StepCodeBench is StepFun's own benchmark, so treat those numbers with some caution.
StepFun also ran two 24-hour agent experiments. In the first, the model tuned an H100 kernel to 508 TFLOPS, edging out Claude Opus 5's 493. In the second, it lifted Qwen3-30B-A3B's AIME24 score from 53.3% to 60% through automated post-training.
The independent check comes from Artificial Analysis, which scored Step 5 Preview at 44 on its Intelligence Index — well above the median of 24 for reasoning models in a similar price tier. Measured throughput was 99.8 tokens per second on StepFun's API.
Pricing: Cheap Per Token, With a Catch
StepFun's API list prices per 1M tokens:
| Token type | Price |
|---|---|
| Input, cache miss | $1.00 |
| Input, cache hit | $0.05 |
| Output, including reasoning | $2.70 |
For comparison, Artificial Analysis puts medians for comparable models at $1.88 input and $10.00 output. That looks like a big win — but there's one caveat. On the index run, Step 5 Preview generated 160M output tokens versus a 92M median. Verbose reasoning chains eat into some of those per-token savings, so real-world task costs may not be quite as dramatic as the sticker prices suggest.
Key Takeaways
- Step 5 Preview is a 600B-total, 27B-active MoE model with a 92-layer, narrow-deep architecture.
- It supports a 1M-token context window with text, image, and video input.
- API pricing is $1.00 input / $2.70 output per 1M tokens, with cheap cache hits at $0.05.
- Artificial Analysis independently scored it at 44 on its Intelligence Index.
- Open weights are scheduled for October 15, 2026, requiring roughly 1.2 TB of BF16 memory.
FAQ
What is Step 5 Preview?
It's StepFun's flagship MoE model designed for agentic coding, professional knowledge work, and finance applications.
Is it open weight?
Not yet. StepFun has scheduled open-weight release for October 15, 2026.
How large is the context window?
1 million tokens, according to StepFun's documentation.
How much does it cost?
$1.00 per 1M input tokens and $2.70 per 1M output tokens via the API.
Bottom line: if your agents run long sessions with lots of tool calls and you're watching API spend, Step 5 Preview offers a genuinely compelling price-to-performance trade-off — just budget for its chattier reasoning style.
Comments (0)