StepFun Step 5 Preview Hits OpenRouter: Kimi K3 Rival at $1

👁 114 reads

Shanghai-based StepFun has launched its new flagship model on OpenRouter. Step 5 Preview arrived on October 8, 2026, under the model ID stepfun/step-5-preview. It costs $1.00 per million input tokens and $2.70 per million output tokens, matching Kimi K3 with a score of 44 on the Artificial Analysis Intelligence Index. The team plans to release the model weights on October 15. We have compiled all the specs, pricing details, and our own test results in our Step 5 guide.

What StepFun Announced

StepFun first unveiled the model on September 20, but until now it was only accessible through the company’s proprietary platforms. Its arrival on OpenRouter means you can call it using the same API key as GPT, Claude, or Gemini, without needing an account registered in China. Key specifications include:

• Mixture of Experts architecture: 600 billion total parameters, with 27 billion active per token;
• 1 million token context window with up to 64,000 output tokens;
• Multimodal input: text, images (up to 60 per request), and video (up to 128 MB);
• Mandatory reasoning with three effort levels: low, medium, and high;
• Available on OpenRouter exclusively via StepFun, running in fp8 precision.

The company positions it as an “agentic workflow” model, emphasizing coding and financial analysis. In self-reported benchmarks, Step 5 scores 67.7 on DeepSWE (compared to 67.5 for Kimi K3 and 66.9 for GLM-5.3) and 66.4 on FrontierFinance.

Where Step 5 Preview Ranks

Artificial Analysis awarded Step 5 Preview a score of 44 on its Intelligence Index, ranking it 40th out of 226 models. The platform currently lists it as proprietary. Once its weights are released, it would tie for third among open-weight models:

StepFun’s strongest selling point is pricing. In charts accompanying the launch, the company highlights a 65% lower per-task cost compared to GLM-5.3 and Kimi K3 in max mode. According to Artificial Analysis, a benchmark task on Step 5 costs an average of $1.03. The model tends to be verbose: across the AA evaluation suite, it generated 160 million tokens against a median of 94 million. Output speed clocks in at around 87 tokens per second.

Our First Benchmark Results

We ran Step 5 through the 13 benchmark tasks in our Testing Lab (we tested it in Polish). It scored 65.4 points, putting it roughly on par with GLM-5.3 (64.8), but well behind Qwen3.8 Max (76.8) and GPT-6 Luna (80.4). The lower score came down to three empty completions: on the low reasoning setting, the model exhausted its full 6,000-8,000 token reasoning budget and returned no final output. When we increased the headroom to 16,000 tokens on a re-run, it fixed 19 out of 22 mistakes. It also scored 46 out of 46 on an introductory Polish trivia quiz. Read the complete breakdown in our report: Step 5: 13-Task Benchmark.

What Next: Open Weights on October 15

StepFun has promised to release the weights on October 15. They have not yet appeared on Hugging Face, and the license remains unconfirmed, though previous releases from the company used Apache 2.0. At 600 billion parameters, the weights alone in BF16 format will require around 1.2 TB of storage, making this a target for enterprise servers rather than consumer GPUs. Once the weights go live, we will document deployment requirements and actual memory footprints.

For developers, Step 5 is currently an affordable, capable option for coding and long-document analysis via API. For non-English generation, other competitive models exist - see our comparison in Step 5 vs Kimi K3, GLM-5.3, Qwen, and DeepSeek.

// read moreStep 5 Preview: Specs, Pricing, Benchmarks and Polish TestState of AIStep 5 vs Kimi K3, GLM-5.3, Qwen and DeepSeek ComparedBenchmarksHY Image 3.5 Preview Review: Pricing, Text, and EditingBenchmarks

Kinetyra