Step today fully opens its new flagship base model Step5Preview, with a clear positioning - a main model designed for real-world Agentic tasks. It aims to solve the classic "Pareto problem" in AI models: the three goals of capability, efficiency, and cost pull against each other. Historically, advancing in one dimension often required sacrificing another; pushing the Pareto frontier is not about eliminating trade-offs, but expanding the entire boundary through new designs.

The approach this time is a sparse MoE architecture. Step5Preview has a total of 600B parameters, but only activates 27B at a time. It natively supports a context window of up to 1 million tokens and can directly process both text and visual inputs. In Artificial Analysis's Intelligence Index, it scored 44 points, making it among the top three global open-source models. Even more impressive, the single-task cost is only one-eighth of Claude Opus5 - giving users either stronger intelligence with the same budget or stronger capabilities at a lower price, paving the way for both options.

image.png

Its hard skills showed off on several high-difficulty benchmarks. On the CLI subset ALE-CLI of Agents' Last Exam, the financial investment research evaluation FrontierFinance, and the cross-domain deep research evaluation DRACO, Step5Preview ranks second only to GPT-6Astra or Claude Opus5, leaving other open-source models behind; GDPval-AA v2 also adopted the latest results as of September 19th, with DeepSWE v1.1 evaluated under the SWE-agent harness at temperature=1.0 and top_p=0.95.

In programming, Step has built its own StepCodeBench, covering 553 independent code repositories, 9 types of tasks, 20 application areas, and 33 programming languages. Step5Preview stands strong in overall success rate and cross-scenario stability, handling real tasks like bug fixing, feature development, refactoring, and environment configuration. In one demonstration, it read a large amount of development documentation for an ESP32 device, transformed an ESP32-S3 development board into a Vibe Coding keyboard that supports Bluetooth buttons and voice input, and during the process, it opened COM serial ports, adjusted cameras, captured device images, simulated mice, and continuously modified and debugged code based on real errors, all in over three hours.

image.png

Long-term tasks truly show its strength. 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). Step5Preview independently modified code, ran the kernel, and measured throughput. If a solution worked but slowed down, it would give up and continue from the best version. After about 22 hours, it pushed the peak performance to 508 TFLOPS, surpassing Claude Opus5's 493 TFLOPS. The second experiment also lasted 24 hours, asking 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 itself, ultimately increasing the accuracy from 53.3% to 60%, matching Opus5 while using fewer annotated tokens.

In terms of front-end and visual capabilities, it doesn't just write web pages but can also 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 book called "The Guangji Railway Travel Guide" to create a journey query, cost calculation, and route playback system, allowing a small train to run along the historical route, turning historical records into interactive products.

Finance is used as a test of comprehensive abilities. Step has set up three internal evaluations around corporate research, a complex 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. The external benchmark FrontierFinance covers six investment scenarios with 220 expert questions and 11543 scoring criteria. Step5Preview delivered strong results in all four.

From Step3.5Flash, Step3.7Flash to Step5Preview, Step has always bet on the same thing: the next stage 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 model, which has recombined "capability - cost - efficiency," is now waiting for developers to use it as their daily mainstay for Agent tasks.