Step 5 Preview Is Now Live on VEGA
StepFun released Step 5 Preview, its new flagship model for agentic work. It runs on a sparse Mixture-of-Experts architecture with 600B total parameters and 27B active per token, supports a 1M-token context window, and accepts image input. On the Artificial Analysis Intelligence Index it scores 44, and StepFun says its cost per task is substantially lower than models at a comparable level of intelligence.
The model is already live on our platform. Step 5 Preview is available on vega.chat in the model picker, and on api.vega.chat, our API for vibe coding and agents. StepFun will release the model with open weights on October 15.
Coding
StepFun built its own evaluation, StepCodeBench, to reflect software engineering as it happens in practice. It spans 553 independent repositories, 9 task categories, 20 application domains, and 33 programming languages. Step 5 Preview scores 49.0% avg@4 there, the best result among the open-weight models on their table, with particular strength in bug repair, feature modification, and refactoring. In reviews by internal and external experts, about 70% of participants judged the model capable of autonomously solving coding tasks of moderately high complexity.
The same breadth extends past software. Given development documentation and user authorization, Step 5 Preview can work with programmable hardware directly, using cameras, COM ports, screenshots, and simulated mouse input as part of the development and debugging loop.
Long-horizon execution
StepFun tested sustained work in two settings with measurable feedback:
- MLA GPU kernel optimization. The model had 24 hours to optimize an MLA kernel from scratch on an NVIDIA H100 (head dimension 512, production shape of batch 1, 64 heads, 8,192 tokens). It implemented changes, ran the kernel, measured throughput, and discarded candidates that regressed. Its best of 4 runs peaked at 508 TFLOPS for forward and backward after roughly 22 hours, ahead of Claude Opus 5 at 493.
- Automated post-training. In another 24-hour run, Step 5 Preview improved a Qwen3-30B-A3B base model on AIME24 through an automated post-training loop, deciding how to use an API annotator and how to revise the data. The resulting model reached 60% on the official test, up from 53.3%, matching Claude Opus 5 with fewer annotator tokens.
There is also a stranger datapoint. Without any game-specific tuning, Step 5 Preview has been playing Pokémon Red for more than 3,000 turns and 6 million tokens. By turn 3,082 it had unlocked Cut, earned three Gym Badges, and defeated Lt. Surge, putting it roughly a third of the way through the main story. The main story takes an estimated 26 hours of human play.
Knowledge work and finance
For a climate study covering 1,000 locations over 25 years, the model coordinated 950 web fetches in a single agent action and assembled 300,000 monthly records across 11 variables, then used the dataset to analyze regional differences in solar seasonality. In a diesel surcharge review it produced a 17-sheet analytical workbook with 11,057 rows, cross-series reconciliation, formulas, and trend models that a reviewer can trace back to source data.
Finance got dedicated attention. StepFun built three internal benchmarks that check whether a model can retrieve and verify timely financial information, turn data and defensible assumptions into reproducible valuations, and carry a full research process from evidence gathering to a finished report. It also evaluated the model on FrontierFinance, an external benchmark covering six investment use cases through 220 expert-written questions and 11,543 evaluation criteria.
Benchmarks
Step 5 Preview ran at the "High" effort setting; comparison models ran at "Max". Benchmarks marked † are internal.
| Benchmark | Step 5 Preview | GLM-5.3 | Kimi K3 | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5 |
|---|---|---|---|---|---|---|
| GPQA Diamond | 93.5% | 91.7% | 93.5% | 96.1% | 93.7% | 93.2% |
| AA-LCR v1.1 | 88.3% | 79.7% | 88.7% | 80.7% | 85.3% | 79.3% |
| DeepSWE v1.1 | 67.7% | 66.9% | 67.5% | 74.1% | 67.4% | 74.0% |
| Terminal-Bench v2.1 | 85.0% | 83.9% | 85.0% | 88.4% | 91.4% | 89.1% |
| CyberGym | 84.7% | 84.5% | 80.0% | — | — | — |
| StepCodeBench† | 49.0% | 40.2% | 43.9% | 61.0% | — | 63.9% |
| GDPval-AA v2.1 | 1566 | 1645 | 1524 | 1542 | 1735 | 1708 |
| BrowseComp | 88.7% | — | 91.2% | 91.5% | — | 90.2% |
| HLE w/ tools‡ | 59.4% | 62.5% | 56.0% | 57.2% | 65.0% | 63.6% |
| FinStepBench-LiveSearch† | 74.5% | 73.3% | 70.9% | 74.5% | — | 76.2% |
| FinStepBench-FinanceDR† | 55.8% | 53.3% | 48.9% | 45.0% | — | 59.1% |
| FrontierFinance | 66.4% | 64.1% | 62.6% | 55.0% | — | 69.7% |
| MMMU-Pro | 76.0% | — (text only) | 81.0% | 87.0% | — | 85.0% |
For HLE w/ tools‡, Step 5 Preview and GLM-5.3 were evaluated on the text-only subset while other models ran on the full dataset, so those results are not directly comparable. The pattern across the table is consistent with the "Pareto frontier" framing in StepFun's announcement: the model rarely tops the chart outright, but it stays close to the leaders at a lower cost per task.
Try it on VEGA
Open vega.chat, pick Step 5 Preview in the model list, and send your first task. If you build with agents or code assistants, connect through api.vega.chat. The full announcement with interactive demos, including the Pokémon replay and the 3D apps, is on stepfun.com.