Step 5 Preview Returns to the Global Top Tier
Author | Wu Haiming
Editor | Li Na TAG62@@Cen Feng
Two years ago, when StepFun first made its public debut in the foundation-model community, it took the stage by storm with Step 1. Later that same year, it released Step 1V and the trillion-parameter foundation model Step 2 in succession. At the time, the industry still believed that the bigger the model, the more capable it was. Whoever pushed the Scaling Law forward would be closer to the center of the table. But over the past year, its exposure has seemed to focus more on the other end: multimodal models, smartphones, cars, on-device models, and agents. By contrast, the flagship foundation model that needs to be compared head-on with Qwen, Kimi, and GLM did not seem to be the center of discussion. Step 5 Preview changed the landscape.
On the Artificial Analysis Intelligence Index, it scored 44 points, ranking second among open-source models globally and placing it in the same league as leading flagship models. This result also proves one thing:
Even as StepFun bets on devices and agents, it remains at the frontier of foundation-model capabilities and has not fallen behind.
But for agents, making the leaderboard only earns an entry ticket. The real challenge is whether, after leaving the benchmark, the model can complete a task from start to finish over dozens or even hundreds of steps. To find out, we conducted hands-on tests.
01
Hands-on test: Building a 3D game
to evaluate Step 5 Preview's capabilities
“Super Mario” made an entire generation believe for the first time that even a simple set of rules could come together to form a complete world. Whenever we pressed the jump key as children, the same question would always come to mind: What if, one day, I could make a game of my own?
At that time, the idea would usually never make it past the threshold, because a game is not an image, a few story sequences, or a piece of code; it is the combined result of characters, scenes, collision detection, interaction, asset export, and project building. It would often take a large team—or several multinational companies working together—to push an idea like that into a runnable system.
Today, for this test, we chose to put the task into a professional software toolchain: Claude Code CLI as the Agent workbench, Blender for 3D assets and levels, and Three.js/Vite for rendering the web game.
The assignment was to have Step 5 Preview create a runnable 3D mini-game called “Star Leap Island: Energy Delivery.” The detailed prompt was as follows:
Code block
Please use Claude Code CLI to invoke the locally installed Blender and complete a runnable 3D mini web game.
- The game must be titled “Star Leap Island: Energy Delivery.” Its protagonist is an original mushroom-headed character named “Gulu.” Do not use Mario, Toad, Nintendo logos, textures, or sound effects.
- In Blender, create the character, six floating platforms, five energy orbs, and one finish beacon; export the .blend and GLB files;
- Then use Three.js/Vite to create the web game, load the GLB assets, and implement movement using WASD/arrow keys, jumping with Space, respawning after falling, a collection counter, a timer, a restart button, and a level-completion prompt.
- Also generate a test script, run npm test and npm run build, and save the source code, build artifacts, logs, and acceptance report to the designated working directory.
What this tests is whether the model can turn an idea into an engineering asset that can be inspected, built, and iterated on. We did not write code or operate Blender on the model's behalf; instead, we broke the original large task into four stages: first complete the protagonist “Gulu,” then build the floating-island level, then export the GLB assets, and finally move on to the web-game project. Within each stage, code writing, Blender operation, file generation, and debugging were still completed by Step 5 Preview.
The game character “Gulu” stands on the game’s starting floating platform
The game’s endpoint floating platform, featuring an energy orb and an endpoint beaconTAG218@@
Complete game design mockup: 6 floating platforms, the game character “Gulu,” and the endpoint beaconTAG223@@After breaking the task down, it began progressing steadily. Step 5 Preview first completed the character and level assets, then integrated the Blender-generated models into the Three.js/Vite project and went on to handle movement, collision detection, energy orb collection, and endpoint detection. There were still revisions and debugging along the way, but ultimately, the project was not torn down and rebuilt. Instead, work continued on the existing results until the project could run.In the end, “Star Leap Island: Energy Delivery” ran properly: players could control “Gulu” to move among the 6 floating platforms, collect 5 energy orbs, and reach the endpoint beacon. The character, environment, collision handling, basic interactions, and web project had all been connected. Although some visual and control details were still rough, it was no longer just a set of static assets, but a complete demo that could actually be played.Come experience this 3D web game, “Star Leap Island: Energy Delivery,” in action:Execution Results:Step 5 Preview can understand complex goals andby invoking engineering tools such as Blenderpush the task through to the final build.TAG248@@Therefore, what this round of testing truly tested was not merely code generation, but whether, in a long-running task spanning Blender, the file system, and the web project, the model could stay focused on the objective, use tools, continue based on intermediate results, and ultimately complete the delivery.To summarize the test results in one sentence:Step 5 Preview now has the end-to-end execution capability to go from natural-language requirements to building a game project,and when faced withTAG267@@complex projects,TAG273@@it canTAG280@@useTAG286@@clearer stage-by-stage breakdownsTAG292@@to complete the task.According to StepFun’s official announcement, Step 5 Preview’s Coding capabilities are not limited to traditional coding tasks. It can handle complex software engineering (including long-horizon planning tasks), frontend and visual development, and programmable hardware scenarios as well; when a task runs for several hours and requires continual adjustment based on the results, the model can keep moving forward. In evaluations involving internal and external experts, about 70% of evaluators believed that Step 5 Preview could autonomously complete coding tasks of medium to high complexity.(Leiphone(WeChat Official Account: Leiphone))Meanwhile, Step 5 Preview also demonstrates superior performance in task scenarios such as professional knowledge work and financial analysis: on ALE-CLI, the CLI subset of Agents' Last Exam, the financial investment research benchmark FrontierFinance, and the cross-domain deep research benchmark DRACO, it ranks second only to GPT-6 Astra or Claude Opus 5, ahead of all other open-source models evaluated.
Step 5 Preview achieved excellent results in information retrieval,valuation, and complete financial research workflows across four financial benchmarksOnce AI agents truly reach production-grade deployment, competition is no longer merely about whether a model can complete a single step, but whether it can execute continuously across documents, tools, and turns while keeping error rates, execution time, and costs within acceptable bounds. Step 5 Preview's 1M-token context and design for long-horizon tasks provide the foundation for this kind of work; at the same time, it delivers reliable production-grade performance across a variety of real-world business scenarios.
02
Why Is Step 5 Preview Worth Discussing?
According to a screenshot of the model page on Artificial Analysis (the AA leaderboard), Step 5 Preview has an Artificial Analysis Intelligence Index score of 44, an output speed of 99.8 tokens/s, and a cost of $0.71 per Intelligence Index task; input is priced at $1.00 per million tokens, output at $2.70 per million tokens, and the cache discount is 95%. The page also shows that Step 5 Preview supports text and image inputs and text output, with a context window of 1M tokens and a parameter count of 600B. This shows that the Step 5 Preview introduced by StepFun is a combination of “intelligence, speed, cost, and context window.”
Screenshot of the Step 5 Preview model page on the AA leaderboardJudging by AA's Intelligence Index, Step 5 Preview has already joined the ranks of frontier models. The index covers multiple categories of tasks, including knowledge work, agentic workflows, terminal use, coding, long-context tasks, and document reasoning. For a model designed for agentic tasks, this is closer to real-world usage than considering any single question-and-answer leaderboard in isolation.Task cost is even more critical. An Agent does not answer a question once and done; it repeatedly reads context, invokes tools, generates code, runs tests, and continues modifying its work based on errors. As tasks grow longer, per-token pricing and caching strategies directly affect whether users are willing to let the model attempt several more rounds. This is also the key factor determining whether developers and enterprise users will choose the model and whether it can achieve truly scaled commercial deployment. As shown in the screenshot below, with a per-task cost of $0.71, Step 5 Preview ranks fourth among 654 models (among the lowest-cost models), equivalent to just 35% of Kimi K3's cost. At the same time, the model pushes the Pareto frontier for the optimal balance of model capability, efficiency, and cost further outward. For long-horizon Agents, this ranking means it is not only competitive on the “capability leaderboard”; it also provides a more compelling answer to the question of whether “multiple rounds of experimentation are worth it.”
03
With three cards in hand,
StepFun moves to the center of the table in China’s domestic foundation-model game
Step 5 Preview’s entry into the global frontier tier on the AA leaderboard marks a landmark in reassessing StepFun. StepFun is combining frontier foundation models, a comprehensive model matrix spanning all modalities, and mass-production deployment on end-user devices into a unified force, shaping its own competitive moat.1. Step 5 Preview puts StepFun in the global frontier-model arena. Particularly for Agent tasks, long context windows and lower invocation costs mean the model can handle longer task chains, larger engineering projects, and more files without having to break complex tasks into too many fragments.2. Over the past quarter, StepFun released Step Edge, the StepAudio 3 series, and then Step 5 Preview. Its model matrix now covers a complete, all-modal lineup spanning cloud flagship models, the efficient Flash model, on-device models, and specialized capabilities such as speech, vision, generation, and GUI interaction. According to official sources, StepAudio 3 Realtime ranked first with a score of 98.9% on the Artificial Analysis Conversational Dynamics leaderboard, with 99.7% accuracy on Speech Reasoning; StepAudio 3 ASR achieved a 1.7% word error rate on the non-streaming speech recognition accuracy leaderboard, tying for first place. Listening, speaking, seeing, generating, and operating are not fragmented capabilities; instead, they are being organized into a complete model system.3. According to publicly available official data, StepFun’s models have entered mass deployment in real-world device scenarios such as smartphones and cars; as of the end of 2025, its models had been installed on more than 42 million phones, serving nearly 20 million user visits per day, and covering more than 50% of leading smartphone brands in China. Compared with pure-API model companies, the value of device entry points lies in high-frequency, real-world, and complex usage: low latency, limited computing resources, noisy environments, privacy constraints, and user habits will continuously put the models to the test in real-world conditions.As the industry evolves, competition among foundation-model companies is no longer merely a comparison of individual model capabilities, but a systemic contest between full-stack technical capabilities and commercial deployment. DeepSeek has the edge in global developer influence, Alibaba’s Qwen is backed by its cloud and ecosystem, Zhipu AI focuses deeply on enterprise delivery, and Kimi has stronger mindshare as a consumer product. StepFun’s unique position, meanwhile, is “AI + devices”: frontier models handle complex reasoning, models spanning all modalities handle perception and expression, and on-device models provide low-latency, high-frequency entry points.At the same time, the path StepFun has chosen is no easy one. The device ecosystem involves smartphone manufacturers, automakers, operating systems, and app permissions; for Agents to be truly deployed, security, privacy, cross-app execution, and user trust must also be addressed. Step 5 Preview demonstrates that StepFun has re-entered the arena of first-tier frontier models. Together with its existing model matrix spanning all modalities and its mass-production deployment capabilities on devices, it directly propels StepFun to the center of the domestic AI table.The endgame of the large-model market remains unsettled, and the lead enjoyed by top vendors will not remain unchanged. Model capabilities, cost structures, device entry points, and workspace experiences could all reshuffle the deck. The significance of Step 5 Preview lies in bringing StepFun back into this game; what really matters next is whether the model can carry tasks through to completion, time and again, across more real-world workflows. Online demo link: https://studio.stepfun.com/
This is an original article from Leiphone. Unauthorized reproduction is prohibited. See theReproduction Noticefor details.
