Step 5 Preview, the new flagship foundation model, is now fully released. Its positioning is clear — a main model designed for Agentic tasks in the real world. It aims to solve the classic "Pareto problem" of AI models: the three goals of capability, efficiency, and cost are constantly conflicting. Historically, moving forward in one dimension often meant sacrificing another. But advancing the Pareto frontier is not about eliminating trade-offs but pushing the entire boundary outward with new designs.

The approach this time is a sparse MoE architecture. Step 5 Preview has a total of 600B parameters, but only 27B are activated each time. It natively supports a context window of up to 1 million tokens and can directly process both text and visual inputs. In the Artificial Analysis Intelligence Index, it scored 44 points, making it among the top three global open-source models. Even more impressive, its single-task cost is only one-eighth that of Claude Opus5 — offering two paths: stronger intelligence with the same budget or stronger capabilities at a lower price. On ALE-CLI, a subset of Agents' Last Exam, the financial investment research evaluation FrontierFinance, and the cross-domain deep research evaluation DRACO, it ranks just behind GPT-6Astra or Claude Opus5, leaving other open-source models in the rear. GDPval-AA v2 also uses the latest results up to September 19th. DeepSWE v1.1 was evaluated under the SWE-agent harness with temperature=1.0 and top_p=0.95.

image.png

In programming, Step built its own StepCodeBench, covering 553 independent code repositories, 9 types of tasks, 20 application areas, and 33 programming languages. Step 5 Preview stands firm in overall success rate and cross-scenario stability, handling real tasks such as bug fixing, feature development, refactoring, and environment configuration. In one demonstration, it read a large amount of documentation on the ESP32 device and transformed an ESP32-S3 development board into a Vibe Coding keyboard supporting Bluetooth buttons and voice input. During the process, it opened COM serial ports by itself, adjusted the camera, captured device images, simulated the mouse, and continuously modified the code to debug based on real errors, completing the task for over three hours.

Long-term tasks show its true value. The first experiment gave it 24 hours and one H100 to optimize a MLA GPU kernel from scratch (head dimension 512, batch 1, 64 attention heads, 8192 tokens). Step 5 Preview independently modified the code, ran the kernel, and measured throughput. When encountering a solution that worked but slowed down, it abandoned it and continued from the best version. After about 22 hours, it reached a peak performance of 508 TFLOPS, surpassing Claude Opus5's 493 TFLOPS. The second experiment also lasted 24 hours, allowing it to improve the performance of a Qwen3-30B-A3B base model on AIME24 through automated post-training. The model decided how to use the annotated API and adjust the data, ultimately increasing the accuracy from 53.3% to 60%, matching Opus5 but using fewer annotated tokens.

In terms of front-end and visuals, it doesn't just write web pages; it can call Blender to build and repeatedly modify 3D assets, then integrate them into Three.js for interactive applications and games. It even extracted information from a 1922 "Guangjiu Railway Travel Guide," creating a journey query, cost calculation, and route playback system, making a small train run along the historical line, turning historical materials into interactive products.

Finance is considered a test of comprehensive ability. Step built three internal evaluations around company research, a high-level process: FinStepBench-LiveSearch tests the ability to retrieve and verify financial information as needs change, CorporateValuation tests converting data and assumptions into reproducible valuations, and DeepResearch tests the entire process from evidence collection to a complete research report. External benchmarks like FrontierFinance cover six investment scenarios with 220 expert questions and 11,543 scoring criteria. In all four, Step 5 Preview delivered strong results.

From Step3.5Flash and Step3.7Flash to Step5Preview, Step has always bet on the same thing: the next phase of scaling is not just about adding more Compute, but using Compute more efficiently. After today's full release, the complete model weights will be released on October 15th. This open-source flagship, which has recombined "capability-cost-efficiency," is now waiting for developers to use it as their daily mainstay for Agent tasks.