StepFun models, access and benchmark guide
StepFun model and access guide with a practical US-English evaluation brief and links to publisher materials.

👁 111 reads
A character-based tariff can define “character” differently from the number your text editor reports. StepAudio’s published counting convention makes that relevant in English. I would retain the provider’s actual billed count for scripts containing punctuation, numbers and abbreviations.
Choosing the first job
Read a short equipment serial with mixed digits and letters; compare source characters, billed units and intelligibility.
Models in the November 17 snapshot
| Model and track | Placement | Reported score |
|---|---|---|
| StepAudio 2.5 TTS (Aug 2026) (AA provider voice) | 16 | 1215 Elo |
| StepAudio 2.5 TTS (AA controlled voice) | 15 | 1076 Elo |
| Step 5 Preview (AA intelligence) | 25 | 29 index points |
Release record. Publisher documentation identifies the product family. The ranked configurations include StepAudio 2.5 TTS (Aug 2026), StepAudio 2.5 TTS, Step 5 Preview.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- https://huggingface.co/stepfun-ai (accessed November 17, 2026)
- https://github.com/stepfun-ai (accessed November 17, 2026)
- https://artificialanalysis.ai/text-to-speech/leaderboard/provider-voice (accessed November 17, 2026)
- https://artificialanalysis.ai/text-to-speech/leaderboard/controlled-voice (accessed November 17, 2026)
- https://artificialanalysis.ai/leaderboards/models (accessed November 17, 2026)
More on this from QUANTSLANT
benchmarks·Stability AI models, access and benchmark guideStability AI model and access guide with a practical US-English evaluation brief and links to publisher materials.