Step 5 Preview, a sparse Mixture-of-Experts model from Shanghai Jieyue Xingchen (StepFun), debuted on September 20, 2026, immediately tying for the top spot on the Artificial Analysis Intelligence Index with a score of 44. The model activates roughly 27 billion parameters per token out of a 600 billion total, supports a one-million-token context window, and accepts both text and vision input. StepFun framed the release as a deliberate strike against Western frontier labs, publishing vendor benchmarks, API pricing, and a hard timeline for open-weight release before most observers had finished downloading the system card.
Step 5 Preview: Architecture and Benchmarks
The Step 5 Preview architecture is a textbook sparse-MoE design. StepFun disclosed 600 billion total parameters with 27 billion active per forward pass, a ratio the company argues yields better inference economics than dense flagship models from Anthropic or OpenAI. The model accepts interleaved text and image inputs and processes up to one million tokens of context, putting it in the same operating envelope as Gemini 2.5 Pro and Claude Opus 5. On the vendor’s own evaluation suite, Step 5 Preview posts a 93.5 percent score on GPQA Diamond, 85.0 percent on Terminal-Bench v2.1, and 88.7 percent on BrowseComp, numbers that StepFun used to justify tying Kimi K3 Max atop the Artificial Analysis Intelligence Index at 44.
Pricing and the Claude Opus 5 Comparison
StepFun priced Step 5 Preview at $1.00 per million input tokens and $2.70 per million output tokens through its public API, with a 95 percent cache discount available for repeated prompt prefixes. The headline claim, however, is task-level cost: StepFun asserts that running a representative software-engineering workload on Step 5 Preview costs roughly one-eighth what the same workload costs on Anthropic’s Claude Opus 5. The figure is a vendor number rather than an independent measurement, but it lands at a moment when enterprise procurement teams are actively shopping for cheaper reasoning models, and the API price undercuts Opus 5 by a wide margin on raw tokens as well.
Engineering Notes: Kernels and Post-Training
StepFun’s technical blog accompanying the launch highlights two pieces of internal work. First, a 24-hour H100 MLA GPU kernel optimization push produced 508 TFLOPS of sustained throughput, compared with the 493 TFLOPS StepFun measured on Claude Opus 5 under identical conditions. Second, a 24-hour automated post-training run on the smaller Qwen3-30B-A3B base lifted AIME24 accuracy from 53.3 percent to 60 percent, evidence that the team’s reinforcement-learning pipeline can deliver double-digit gains in well under a day of compute. Both claims are vendor-reported and have not yet been reproduced externally.
Company Background and a Skipped Generation
StepFun was founded in April 2023 by Jiang Daxin, a former vice president at Microsoft Research Asia, and has since grown into one of China’s most prolific model labs. The company shipped Step-2, Step-3, and Step-3.7-Flash in successive waves before jumping directly to Step 5 Preview, skipping the entire Step-4.x line that competitors and customers had been expecting. StepFun has not publicly explained the naming gap, though engineers familiar with the lab’s roadmap suggest the Step-4 designation was reserved for an internal research project that did not graduate to a commercial release. The jump mirrors a pattern at other Chinese labs, where internal codenames and external version numbers are kept deliberately far apart.
Open Weights and Industry Implications
StepFun says full open weights for Step 5 Preview will arrive on October 15, 2026, via a Hugging Face repository that, as of launch day, remains empty except for a placeholder README. The company is targeting AI coding assistants, software-engineering agents, financial modeling, and broader professional knowledge work as the primary deployment cases. The release fits a now-familiar 2026 rhythm for Chinese frontier labs: publish a closed-preview model that briefly claims a Western-built leaderboard, generate several days of comparative coverage, then drop weights weeks later once the conversation has moved on. Step 5 Preview is the clearest example yet of that cadence in action, and whether its open release can sustain the early benchmark hype will be the story to watch through the rest of the year. Step 5 Preview, in short, has arrived as both a technical statement and a marketing one, and the industry will measure the gap between the two once the weights finally land.
Step 5 Preview continues to define the competitive landscape, and the developments this week underscore how quickly the underlying economics and market structure are shifting. Analysts expect the next month to bring additional disclosure around partnerships, customer commitments, and benchmark performance, all of which will shape how enterprises and consumers evaluate the trade-off between cost, capability, and reliability. The early signal points to a market in which step 5 preview sets the new baseline rather than the ceiling, with rivals forced to match on performance or price to remain relevant.
Step 5 Preview has become the defining story of the AI beat this week, and the broader industry is responding. Analysts at major research desks have begun updating their forecasts, and customer commitments are likely to follow within days. The pace of step 5 preview’s rollout sets a benchmark that competitors will struggle to match on cost without sacrificing capability, and capability without sacrificing cost. Expect additional disclosures over the coming weeks as partners and customers publish their own evaluations and reference deployments.