DeepSeek V3 (Dec '24) vs Step 5 Preview
Which is better for the money: DeepSeek V3 (Dec '24) by DeepSeek or Step 5 Preview by StepFun? Intelligence, total benchmark cost, and end-to-end time from Artificial Analysis & other sources, snapshot 2026-10-24.
Trade-off
Verdict
Depends on your budget
Step 5 Preview scores 35.2 points higher (43.7 vs 8.5) but costs 47× as much to run the benchmark ($1190 vs $25.4; premium vs budget cost classes). They are not alternatives at one budget: pick Step 5 Preview when you need intelligence above 8.5, and DeepSeek V3 (Dec '24) when 8.5 is enough.
| Metric | DeepSeek V3 (Dec '24) | Step 5 Preview | Difference |
|---|---|---|---|
| Intelligence (higher is better) | 8.5 | 43.7 | +414.1% Step 5 Preview is 5.1× more intelligent |
| Total AA cost (lower is better) | $25.37 | $1,190.26 | +4591.6% Step 5 Preview is 47× more expensive |
| End-to-end time (lower is better) | 11.89s | 12.06s | +1.4% Step 5 Preview is 1.4% slower |
| Price per 1M tokens (in / out) | $0.32 / $0.89 | $1.00 / $2.70 | |
| Input modalities | Text | Text, Image |
Step 5 Preview vs DeepSeek V3 (Dec '24): 414.1% more intelligence, 4591.6% more cost to run the AA benchmark, and 1.4% slower end-to-end.
Today's Pareto frontier
Against the frontier
- DeepSeek V3 (Dec '24) is off the frontier: at the same or lower cost, GPT-6 Luna (Low) scores 21.50 for $10.6.
- Step 5 Preview is off the frontier: at the same or lower cost, GPT-6.1 Sol (Max) scores 51.80 for $1082.
DeepSeek V3 (Dec '24)Step 5 PreviewToday's Pareto frontierOpen-weights frontierEvery other model
Matchup
Compare something else
Agents:
Accept: text/markdown or ?format=md returns this page as Markdown. JSON: /api/v1/compare/deepseek-v3/step-5 · Markdown: /api/v1/compare/deepseek-v3/step-5?format=md · llms.txt · agent skill