Step 5 Preview
Step 5
StepFun's jump towards the frontier: strong benchmarks, a million tokens, and open weights promised.
Updated
The read
Step 5 Preview is a sparse model of about 600 billion parameters with 27 billion active, a million-token window and image and video input, released in September. It scores level with Kimi K3 on the Artificial Analysis index for a third of the input price, though human voters rank it further down. StepFun has promised the weights for 15 October; until they appear and the licence is known, treat it as a hosted preview.
Where it shines
- Level with Kimi K3 on the Artificial Analysis index
- Million-token context with image and video input
- A dollar in and $2.70 out per million tokens
Worth knowing first
- Still a preview, with the weights and licence not yet published
- Human voters rank it well below its benchmark scores
Index score
83.2
Provisional. This score is read from the product's public capability, not from the one-prompt rebuild the fully reviewed entries went through, so treat it as a placing rather than a verdict.
- Rank
- 29 of 73
- Builds
- LLMs
- Category
- Frontier
- Output
- Text, images and video in, text out
- Pricing
- Paid only
- Best for
- Long-context multimodal work at a mid-tier price
Scores are our own editorial judgement, weighted craft 55, speed 10, control 15 and value 20.
Also worth a look
Full index- 01Claude Opus 5.5Claude Opus 5.5ProvisionalFable-level answers for less than the Opus it replaces. The new default.Frontier modelPaid only92.8
- 03Gemini 4 ArgonGemini 4 ArgonProvisionalTop of the human-preference leaderboard, and almost nobody can use it yet.Frontier modelPaid only90.6
- 04GPT-6.1 SolGPT-6.1 SolProvisionalNear-Astra answers at a fifth of the price. The OpenAI model most teams should deploy now.Frontier modelPaid only90.3
