Model Card | Step 5 Preview

🌏 中文版

Model Info

Capability Highlights

• Scores 44 on Artificial Analysis's independent Intelligence Index — on par with Kimi K3 and Grok 4.6 (high), just one point behind GLM-5.3 — while costing only $0.72 per task, about 42% less than Gemini 3.8 Flash (high, $1.24)
• Uses a 92-layer narrow-deep Transformer rather than a wider architecture, paired with MTP-3 speculative decoding, FP8 MoE, and KV-cache offload; StepFun reports more than a 3x end-to-end speedup in long-horizon RL training
• In a 24-hour long-horizon test, it optimized an MLA GPU kernel on an H100 to 508 TFLOPS, beating Claude Opus 5's 493 TFLOPS; in a separate test it lifted Qwen3-30B-A3B's AIME24 score from 53.3% to 60% through automated post-training, matching Claude Opus 5 while using fewer annotator tokens
• Coordinated 950 web fetches within a single agent action to assemble a climate-research dataset spanning 1,000 locations and 25 years (300,000 monthly records across 11 variables)

Benchmark Results

⚠️ Except for the Artificial Analysis Intelligence Index, every Step 5 Preview figure above is self-reported by StepFun, run at high reasoning effort while Claude Opus 5 / GPT-6 Astra ran at their respective max effort — not a strictly apples-to-apples comparison — and independent reproduction is still pending. The Artificial Analysis Intelligence Index is an independent third-party evaluation that also measured API output speed at about 99.8 tokens/second.

Versus Predecessor / Competitors

Step 5 Preview is StepFun's first model under the "Step 5" label; its predecessors, Step 3.7 Flash and Step 3.5 Flash, sit in a different, mid-tier class, so there's no strict apples-to-apples predecessor score in the table above. Against same-tier competitors, Step 5 Preview trails Claude Opus 5 and GPT-6 Astra across coding and finance benchmarks (DeepSWE, StepCodeBench, FrontierFinance, DRACO) by roughly 5-14 percentage points, but on the broader Artificial Analysis Intelligence Index it's just one point behind GLM-5.3 and ties Kimi K3 and Grok 4.6.

The real differentiator is pricing strategy: StepFun is positioning Step 5 Preview as a "Pareto frontier" play — trading a bit of peak intelligence for a dramatically lower cost. Artificial Analysis estimates a per-task cost of just $0.72, under a tenth of Claude Fable 5.1's $7.63, and 42% below Gemini 3.8 Flash's $1.24. The trade-off is verbosity: the same evaluation run produced 160 million output tokens versus a 92-million median, eating into some of the per-token savings, and output speed (around 100 tokens/second) trails Gemini 3.8 Flash's 331 tokens/second by a wide margin.

Implications for Agent Development

The 1M-token context window combined with native text/image/video input suits agent tasks that accumulate long-running context — StepFun's own climate-research demo coordinated 950 web fetches within a single agent action while maintaining state throughout. For tasks that need "sustained progress plus repeated tool calls," a long context directly reduces the need for an external memory or RAG layer.

• If you're building a long-horizon research agent (deep research, multi-round data gathering and synthesis): Step 5 Preview's 1M context and its ability to coordinate hundreds of tool calls within a single action make it a candidate for replacing complex "chunked-retrieval-plus-stitching" pipelines, especially when the task can tolerate somewhat higher latency and verbosity
• If you're building a financial-analysis agent: the FrontierFinance and FinStepBench results show StepFun specifically optimized for finance, with valuation, due-diligence, and deep-research tasks landing close to Claude Opus 5 at a fraction of the API cost
• Not a fit: real-time or low-latency interaction (output speed is only around 100 tokens/second, far slower than Gemini 3.8 Flash's 331 tokens/second); also not a fit if you need to self-host today, since open weights aren't due until 2026-10-15 — for now Step 5 Preview is only accessible through StepFun's API

Today's Takeaway

Step 5 Preview's "narrow-deep" architecture choice — 92 layers, no widening — is a notable outlier. Most recent flagship-model races have leaned toward widening layers or adding more experts, while StepFun instead stacked more layers to gain longer multi-hop reasoning paths during long prefill. It's a reminder that when everyone is chasing the same benchmark leaderboard, the architectural path used to reach a given score can matter more than the score itself — because that path determines whether further gains are possible within the same cost structure.

References

• StepFun official announcement: Step 5 Preview — Advancing the Pareto Frontier
• StepFun official docs: Step 5 Preview model specs
• MarkTechPost: StepFun Launches Step 5 Preview: A 600B-Total, 27B-Active MoE Model With 1M Context for Long-Horizon Agentic Work
• OfficeChai: China's StepFun Releases Step 5 Preview, Beats Gemini 3.8 Flash On Performance And Cost
• Artificial Analysis: Step 5 Preview model evaluation page