StepFun Step 5 Preview: 600B Sparse MoE Agent Model with Claude Code Support
StepFun launched Step 5 Preview, a 600B sparse Mixture-of-Experts model activating 27B parameters per token with a 1M token context window. At $1.00 input and $2.70 output per million tokens, it undercuts rival pricing while offering native Claude Code integration.
Impact: Medium
Why it matters
You can cut agent execution token spend by replacing frontier models with a $1/$2.70 per 1M token endpoint integrated directly into Claude Code.
TL;DR
- 01Step 5 Preview activates 27B of its 600B parameters per token across a 1M context window.
- 02Priced at $1.00 input and $2.70 output per 1M tokens, it significantly undercuts standard frontier API rates.
- 03Verbose reasoning output (160M tokens generated vs 92M benchmark median) requires output token monitoring.
Key facts
- Total / Active Parameters
- 600B total / 27B active per token
- Context Window
- 1,000,000 tokens
- API Input Pricing
- $1.00 per 1M tokens
- API Output Pricing
- $2.70 per 1M tokens
- Artificial Analysis Intelligence Index
- 44 (vs 24 price tier median)
- DeepSWE v1.1 Benchmark
- 67.7 (self-reported)
Model Topology and Inference Characteristics
Step 5 Preview is a sparse Mixture-of-Experts (MoE) network comprising 600B total parameters, with 27B activated parameters per token (~4.5% active weight footprint). Built over a 1M-token context window, the model uses a 92-layer deep stack designed specifically to optimize implicit multi-hop agent reasoning during extended prefill cycles. Inference incorporates MTP-3 speculative decoding, FP8 MoE quantization, and KV-cache offload.
Pricing and Benchmark Results
API list pricing challenges established frontier models:
- Input Price: $1.00 per 1M tokens (industry median: $1.88)
- Output Price: $2.70 per 1M tokens (industry median: $10.00)
- Independent Intelligence Index (Artificial Analysis): 44 (median for comparable tier: 24)
- Output Speed: 99.8 tokens per second
On coding benchmarks, StepFun reports self-reported scores of 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench, and 80.5 on ProgramBench, trailing top models like GPT-6 Astra and Claude Opus 5. In agentic tuning tasks, it tuned an H100 kernel to 508 TFLOPS (compared to 493 for Claude Opus 5). Note that its high verbosity (generating 160M tokens vs. a 92M median on benchmark runs) partially reduces the raw per-token cost advantage.
Try it in 2 minutes
# Example API payload for StepFun reasoning model with high effort
curl https://api.stepfun.com/v1/chat/completions \
-H "Authorization: Bearer $STEPFUN_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "step-5-preview",
"reasoning_effort": "high",
"messages": [{"role": "user", "content": "Analyze repository AST bottlenecks"}]
}'bash
✓ When to use
- When running high-volume multi-agent search loops that consume millions of input context tokens per run.
- When cost optimization in Claude Code workflows is paramount and a Pareto-efficient frontier alternative is required.
✕ When NOT to use
- When your workloads require absolute peak reasoning accuracy on coding benchmarks, where Claude Opus 5 or GPT-6 Astra hold measurable leads.
- When deploying on edge devices or single-GPU servers prior to the availability of distilled or quantized open weights.
What to do today
- Configure Step Plan in Claude Code to test Step 5 Preview for cost-constrained long-context operations.
- Implement max_tokens bounds on API calls to mitigate verbose reasoning token consumption.
- Prepare hardware budgets (~1.2 TB storage in BF16) ahead of the planned October 15, 2026 open-weights release.
Sources