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 Launches Step 5 Preview With 600B Parameters and 1M-Token Context

StepFun has unveiled Step 5 Preview, a new flagship model built for agentic workloads spanning software engineering, professional knowledge work, and finance. The company's central claim is efficiency: comparable intelligence to leading rivals at a significantly lower cost per task—a positioning reflected in the "Pareto frontier" language of the launch. The model is available now through a hosted API and the StepFun platform, with open weights promised for October 15, 2026.

Model Architecture and Capabilities

Step 5 Preview is a sparse Mixture-of-Experts (MoE) design with roughly 600B total parameters, of which about 27B are activated per token—approximately 4.5% of the total weights. According to Pandaily, StepFun chose a narrow-and-deep layout with 92 Transformer layers rather than a wider network. The team argues that deeper stacks create longer pathways for implicit multi-hop reasoning, which becomes important during extended prefill phases when agents search the web, execute code, and parse tool outputs.

Training relied on on-policy, long-horizon reinforcement learning. StepFun highlights bit-wise alignment between training and inference across MoE routing, alongside MTP-3 speculative decoding, FP8 MoE, and KV-cache offload. The company reports over 3x end-to-end speedup for long-horizon RL runs.

Official specifications include a 1M-token context window, multimodal input (text, images, and video), text-only output, three reasoning effort levels (low, medium, and high), and support for streaming, tool calling, JSON Mode, JSON Schema, and prompt caching. StepFun also notes the model coordinated 950 web fetches in a single agent action and works with Claude Code through its Step Plan integration.

Benchmarks and Independent Scores

In company-reported results via RuntimeWire, Step 5 Preview ran at High effort while competitors ran at Max. It scored 66.4 on FrontierFinance versus Claude Opus 5's 69.7 and GPT-6 Astra's 55, and 83.3 on DRACO against 87.6 and 76.8 respectively. On coding benchmarks, StepFun reports 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench (its own benchmark), and 80.5 on ProgramBench, where both rivals lead on all three.

Two 24-hour agent experiments also stood out: the model tuned an H100 kernel to 508 TFLOPS, beating Claude Opus 5's 493, and lifted Qwen3-30B-A3B's AIME24 score from 53.3% to 60% via automated post-training.

Independently, Artificial Analysis scored the model at 44 on its Intelligence Index—well above the median of 24 for reasoning models in a similar price tier—and measured output speed at 99.8 tokens per second.

Pricing and the Verbose Trade-Off

StepFun's API pricing is $1.00 per 1M input tokens on cache miss, $0.05 on cache hit, and $2.70 per 1M output tokens including reasoning. Artificial Analysis pegs the medians for comparable models at $1.88 input and $10.00 output. However, there is a caveat: the model generated 160M output tokens on the index run compared with a 92M median, so verbose reasoning erodes some of the per-token savings.

Deployment Considerations

Self-hosting is not yet possible, but open weights arrive October 15, 2026. Rough math suggests 600B parameters require about 1.2 TB in BF16 before KV cache, so prospective self-hosters should plan for multi-GPU server hardware.

Key Takeaways

• 600B total parameters, 27B active per token, in a 92-layer deep MoE design
• 1M-token context accepting text, images, and video
• API pricing of $1.00 input and $2.70 output per 1M tokens
• Artificial Analysis Intelligence Index score of 44
• Open weights scheduled for October 15, 2026

The next milestone to watch is whether the promised open weights ship on schedule—and whether independent benchmark runs confirm the cost-efficiency claims once verbose reasoning overhead is factored in.

Meta description: StepFun's Step 5 Preview is a 600B MoE model with 1M context, aggressive pricing, and open weights due October 15, 2026.

Tags: StepFun, Step 5 Preview, Mixture-of-Experts, agentic AI, LLM pricing

Featured image: Abstract visualization of interconnected neural network nodes in deep blue tones, no text or logos.