Beijing-based AI research company StepFun launched Step 5 Preview on September 20, 2026, unveiling one of the most capable sparse Mixture-of-Experts (MoE) models yet available via API — with open weights promised to the developer community on October 15. The announcement positions StepFun as a credible challenger to the frontier models from Anthropic, OpenAI, and Google, offering near-flagship performance at a fraction of the price.
What Was Announced
Step 5 Preview is a sparse MoE architecture with 600 billion total parameters, of which only 27 billion activate per token — roughly 4.5% of the full parameter count. This design allows the model to match or approach the quality of much larger dense models while keeping inference costs low. StepFun has priced API access at $1.00 per million input tokens and $2.70 per million output tokens, with cached-input pricing dropping to $0.05 per million tokens. At those rates, Step 5 Preview’s output costs approximately 18% of Kimi K3’s $15.00-per-million output price.
The model supports a one-million-token context window and up to 64,000 tokens of output per call, accepting text, image, and video inputs. StepFun also introduced three reasoning effort levels — giving developers the ability to trade speed for depth on a per-call basis, a feature that has become increasingly standard among frontier model providers in 2026.
API access launched on September 20, 2026 via StepFun’s developer platform at platform.stepfun.com. Weights are confirmed for public release on October 15, which would make Step 5 one of the most capable open-weight models available for self-hosted deployment.
The announcement accompanies a series of long-horizon benchmark demonstrations that StepFun used to characterize the model’s strengths in sustained autonomous work.
Technical Details
Step 5 Preview uses a 92-layer narrow-deep Transformer layout. The sparse MoE design routes each token through only a small subset of the model’s expert modules, dramatically reducing the compute required per inference call compared to a dense 600B-parameter model. The architecture is particularly well-suited for deployments where long context and agentic reasoning are more important than raw speed on short queries.
On coding and agent benchmarks, Step 5 Preview scores 67.7% on DeepSWE v1.1 — a benchmark measuring a model’s ability to resolve real-world software engineering tasks autonomously. Claude Opus 5 scores 74.0% on the same benchmark. On Terminal-Bench v4, which measures agentic command-line competence, Step 5 Preview reaches 33.3% versus GPT-6 Astra’s 57.9%. StepFun noted these comparisons use Step 5’s High reasoning mode against rivals’ Max modes, which affects direct comparison.
StepFun demonstrated the model’s long-horizon capability through three experiments: a 24-hour autonomous GPU kernel optimization run that reached 508 TFLOPS, a 24-hour automated training improvement loop, and a 3,000-turn Pokémon Red playthrough. These tests suggest Step 5 is designed primarily for tasks that run for hours or days rather than minutes, placing it firmly in the long-horizon agentic AI category.
Industry Impact and Reactions
The release puts StepFun in direct competition with Z.ai’s GLM-5.3, Moonshot AI’s Kimi K3, and frontier models from Anthropic, OpenAI, and Google. On the Artificial Analysis Intelligence Index, Step 5 Preview scores 44, essentially tied with GLM-5.3 at 45 — making it competitive with the best Chinese-origin models currently available. At its price point and with open weights coming, it gives engineering teams a credible alternative to proprietary APIs for agentic coding and long-document work.
The open weights commitment is strategically significant. In recent months, Meta’s LLaMA series and Mistral’s open releases have shown that open-weight frontier models can reshape market dynamics by enabling on-premise deployments, fine-tuning, and cost structures that closed-API models cannot match. Step 5 entering this space on October 15 will give self-hosting teams a new 600B-parameter option for the first time.
The pricing structure is also notable. By offering cached-input pricing at $0.05 per million tokens — a 95% discount over standard input pricing — StepFun is clearly targeting high-volume agentic workflows where the same context is repeatedly referenced across a long session. This positions Step 5 as an economically attractive backbone for enterprise AI agents that maintain large knowledge bases in-context.
What Comes Next
The primary near-term milestone is the October 15 weights release. If StepFun delivers on schedule, it will mark Step 5 as among the first models at this scale to go fully open in 2026. Independent evaluations and fine-tuned variants should begin appearing within days of the public release. The developer community will likely test it against GLM-5.3, Mistral Large, and Meta’s Llama 4 series.
StepFun has not yet released a training-data or architecture paper alongside the Preview launch, which is common practice for preview releases ahead of full open-weight disclosure. Additional technical detail is expected to accompany the October 15 release.
Conclusion
StepFun’s Step 5 Preview is a technically serious release: near-frontier benchmark scores, a genuinely affordable API, and an open-weights commitment that sets a concrete date for community access. For teams building long-horizon agents, coding assistants, or large-document applications, it is worth evaluating against current leaders when weights arrive in October. The model reinforces a broader 2026 trend: capable open-weight models are no longer a compromise — they are becoming the default starting point for serious AI development.
Stay updated on the latest AI news at Evolve Digital.