StepFun Step 5 Preview Surfaces in Third-Party Benchmarks Ahead of Official Launch

Step 5 Preview is a proprietary multimodal reasoning model from StepFun that supports text and image inputs with a 1M token context window.

The Step 5 Preview model from StepFun has surfaced through evaluations on independent benchmark platforms, offering initial public metrics on its capabilities well before any official company announcement. The model was tested via the StepFun API under the identifier step-5-preview-b, allowing third parties to record standardized performance data. This early visibility provides the AI community with concrete numbers on intelligence and speed that can inform comparisons with other frontier models. The timing of these reports aligns with a period of rapid iteration among multiple AI developers seeking to establish positions in the competitive landscape.

What background context exists around StepFun model development?

StepFun operates as a Chinese AI laboratory focused on advanced reasoning systems, and Step 5 Preview represents the latest in its sequence of proprietary offerings. The preview designation indicates an intermediate stage where the model is accessible for testing but not yet positioned for broad commercial use. Data from multiple sources confirm that the company maintains closed-source weights, consistent with industry norms for high-parameter frontier systems. The 1M token context window stands out as a core architectural choice that supports extended input sequences without truncation of earlier context.

Observations as of September 19, 2026, show that Step 5 Preview remains absent from StepFun's official model list, pricing documentation, and developer resources. Paths that would normally reference the model return 404 errors, reinforcing its non-public status at the time of evaluation. This pattern allows laboratories to collect benchmark feedback while retaining control over the final release narrative. The September 18, 2026 release date recorded by evaluators marks the point at which the model became available for external testing.

What technical specifications define Step 5 Preview?

The model accepts text and image inputs while generating text outputs, enabling multimodal workflows that combine visual and linguistic processing. Its 1M token context window permits retention of substantial prior information during extended sessions, which is relevant for tasks involving long documents or multi-turn reasoning. Evaluations treat the model as closed source, meaning internal weights and training details are not disclosed. The combination of these attributes positions the system for applications requiring both visual understanding and extended textual coherence.

How do reported performance metrics position the model?

Benchmark results place the intelligence score at 43.6 on the Artificial Analysis Intelligence Index, a figure that allows direct comparison with other evaluated systems. Output throughput reaches 93 tokens per second, a speed metric that affects responsiveness in production environments. Pricing is listed at $1.00 per million input tokens and $2.70 per million output tokens, establishing a cost structure that can be weighed against performance gains. These values originate from the dedicated Artificial Analysis model page and the Opper AI summary of the same evaluation run.

Benchmark metrics for Step 5 Preview from third-party sources

Metric	Value	Source
Intelligence Index	43.6	Opper AI
Output Speed	93 tok/s	Opper AI
Input Price per 1M tokens	$1.00	Artificial Analysis
Output Price per 1M tokens	$2.70	Artificial Analysis
Context Window	1M tokens	Artificial Analysis
Release Date	September 18, 2026	Opper AI

Why has Step 5 Preview not appeared on official channels?

DataLearner documentation states that as of September 19, 2026, the model is not present on StepFun's public platform, pricing page, or any official documentation. This absence indicates the preview remains in a controlled testing phase rather than general availability. Access through the API for benchmark purposes does not equate to a full product launch, allowing the company to iterate based on external measurements. The 404 errors on documentation paths further confirm that public-facing materials have not yet been updated.

Such staggered releases are a standard mechanism for laboratories to validate model behavior under real query loads before committing to production infrastructure. The early benchmark data therefore serves as an advance signal rather than a complete product announcement. Stakeholders can use the published numbers to prepare integration plans while awaiting formal documentation updates.

What market and stakeholder implications arise from the preview?

The combination of a 43.6 intelligence score, 93 tokens per second throughput, and the stated pricing creates a data point that other developers may reference when planning model selections. Multimodal input support broadens potential use cases beyond pure text tasks. Closed-source distribution limits fine-tuning options but provides a managed API experience that reduces operational overhead for users. The preview status itself may accelerate competitive responses from other laboratories monitoring similar release patterns.

Enterprises evaluating cost structures will note the per-million-token rates as a concrete factor in total ownership calculations. The large context window supports workflows that previously required chunking or summarization steps. Overall, the metrics supply a reference point for assessing value in the current frontier-model pricing environment without requiring an official launch.

What timeline observations have been recorded so far?

1. September 18, 2026: Model release date noted in benchmarks by Opper AI.
2. September 19, 2026: DataLearner observes absence from official StepFun platforms.
3. Ongoing: Evaluations conducted via StepFun API by Artificial Analysis.
4. Future: Potential official launch and updates to documentation.

What expert or platform reactions address the preview data?

Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index. Pricing for Step 5 Preview is $1.00 per 1M input tokens and $2.70 per 1M output tokens. Released September 2026. Created by StepFun. 1M tokens context window. Supports text and image input, outputs text.Artificial Analysis

The Artificial Analysis summary supplies a consolidated view of intelligence, pricing, and capability attributes that analysts can cite when discussing the model. This information forms the basis for any initial positioning statements ahead of wider availability. The explicit mention of the September 2026 release and creator attribution anchors the data to a verifiable origin.

What developments are anticipated next for StepFun models?

An official listing on the StepFun platform and corresponding documentation updates are the logical next milestones once the preview phase concludes. Additional performance reports or integration examples may accompany the full release to illustrate practical applications. The company may also adjust pricing or capabilities based on feedback gathered during the benchmark period. Continued monitoring of the Artificial Analysis and Opper AI pages will reveal any subsequent revisions to the published metrics.

The current third-party data set provides a stable reference until official channels are updated. This sequence of preview evaluation followed by formal release reflects standard industry practice for frontier-model introductions. Observers can track changes in listed scores or pricing as the model moves from preview to production status.