vs
StepFun vs OpenScale
StepFun ships its own multimodal models, many under Apache 2.0. OpenScale resells DeepSeek V4.1 Flash at promotional prices for coding agents.
By The Subconscious Team · Updated
StepFun vs OpenScale: key differences
StepFun is a Shanghai lab with its own models. Step 5 Preview, its agentic flagship, has 1M context and image input, and the cheaper Step 3.7 Flash runs at about 128 tokens per second for $0.20 in and $1.15 out. Key releases ship under Apache 2.0, and first-party inference is hosted in China. OpenScale, a Y Combinator-backed API out of Singapore, serves DeepSeek V4.1 Flash at $0.06 in, $0.003 cached and $0.24 out per million tokens under a 50% promotion, at 200+ tokens per second, with a fast tier at about twice the speed.
StepFun is the pick for vision and video understanding, or for teams that want permissive weights to self-host. OpenScale is the pick for cheap text inference in Claude Code, Codex and similar agents, with zero data retention on Chat Completions and Messages by default.
What StepFun and OpenScale do
StepFun
StepFun is a Shanghai AI lab known for efficient multimodal models, with a mix of proprietary API models and open-weight releases. In September 2026 it launched Step 5 Preview, its new flagship for agentic work: a 600B-total, 27B-active mixture-of-experts model with 1M context and text and image input, which Artificial Analysis scores level with Kimi K3. Its cheaper workhorse, Step 3.7 Flash, came out in May 2026 as a 198B mixture-of-experts vision-language model with only 11B active parameters. It has 256K context, selectable reasoning levels, tool use and structured outputs, and it ships under Apache 2.0. StepFun's own API prices it at $0.20 in and $1.15 out per million tokens, and OpenRouter carries it too.
Example models: Step 5 Preview, Step 3.7 Flash
Full StepFun profileOpenScale
OpenScale is a Y Combinator-backed inference API built by LatentSpace and run out of Singapore. It sells open models behind one endpoint that speaks Chat Completions, the OpenAI Responses API and Anthropic Messages, so Claude Code, Codex, OpenCode, Cline, GitHub Copilot, Zed and other coding agents can point at it with a base URL and key. Its public catalog today is DeepSeek V4.1 Flash in two versions, standard and fast, both with a 1M token context window and up to 384K output tokens.
Example models: DeepSeek V4.1 Flash
Full OpenScale profileShould you choose StepFun or OpenScale?
StepFun vs OpenScale at a glance
| Attribute | StepFun | OpenScale |
|---|---|---|
| Model access | Open (Apache 2.0) and API models | Open weights |
| Flagship models | Step 5 Preview, Step 3.7 Flash | DeepSeek V4.1 Flash |
| Speed | ~128 tok/s on Step 3.7 Flash | 200+ tok/s; fast tier ~2x |
| Price | $0.20 in, $1.15 out (Step 3.7 Flash) | $0.06 in, $0.24 out (promo, DeepSeek V4.1 Flash) |
| Customization | Open weights to fine-tune | N/A |
| Deployment | First-party API, OpenRouter | Serverless API |
| Long context | Up to 1M (Step 5 Preview); 256K on Step 3.7 Flash | 1M context |
Frequently asked questions
What is the difference between StepFun and OpenScale?
StepFun ships its own multimodal models, many under Apache 2.0. OpenScale resells DeepSeek V4.1 Flash at promotional prices for coding agents.
When should I choose StepFun over OpenScale?
Image and video understanding in cost-sensitive agents; Apache 2.0 weights to self-host; StepFun's own Step 5 models.
When should I choose OpenScale over StepFun?
Low per-token prices on DeepSeek V4.1 Flash; Coding agents outside China-hosted inference; Anthropic and OpenAI formats on one key.
Is StepFun or OpenScale cheaper?
StepFun: $0.20 in, $1.15 out (Step 3.7 Flash). OpenScale: $0.06 in, $0.24 out (promo, DeepSeek V4.1 Flash). The cheaper choice depends on the model and workload.
Which has more context, StepFun or OpenScale?
StepFun: Up to 1M (Step 5 Preview); 256K on Step 3.7 Flash. OpenScale: 1M context.
Related comparisons
Running long-horizon agents?
If your agents run past 200K tokens, compare both against Subconscious. Our inference stack treats a long-horizon trace as the primary workload, so speed, cost, and accuracy hold up deep into the trace.