DeepSeek V3
Language Retired Full comparison ↗ deepseek/deepseek-v3 · by DeepSeek
· mixture-of-experts
Pricing — 1 offering(s)
About this offering: Retired on DeepSeek's own API: DeepSeek V3 (deepseek-chat-v3-0324) was replaced by DeepSeek V3.2 on the deepseek-chat endpoint on 2025-12-01, so the prices shown end there. It was an open-weights mixture-of-experts model with a 164K-token context, and the weights remain available for self-hosting.
Input tokens
Retired — no longer sold · last price $0.20 / 1M tokens (input), until 2025-11-30
- $0.20 / 1M tokens (input) Historical 2025-03-24 → 2025-11-30
Output tokens
Retired — no longer sold · last price $0.77 / 1M tokens (output), until 2025-11-30
- $0.77 / 1M tokens (output) Historical 2025-03-24 → 2025-11-30
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how DeepSeek V3 fits into a cost-aware routing setup
See how →Capability profile
How DeepSeek V3 rates across core capability dimensions, with the task-level evidence behind each rating.
Strong on benchmarks; competitive with frontier closed models at a fraction of the cost.
SWE-bench Verified (500 human-verified GitHub issues; patch must pass the repo's own tests; no file-location hints): resolves 42% — fails roughly 3 issues in 5. From the DeepSeek-V3 technical report (V3-1226 checkpoint; agentless framework).
164K context.
Coding benchmarks
View the full coding benchmark leaderboard →| Benchmark | Score | Harness | Source |
|---|---|---|---|
| livecodebench ↗ | 49.2% 2024-12 | — | 📊 leaderboard ↗ |
| aider-polyglot ↗ | 55.1% 2025-03 | aider | 📊 leaderboard ↗ |
| swe-bench-verified ↗ | 42.0% 2024-12 | agentless | 📄 paper ↗ |
| bigcodebench ↗ | 27.7% 2025-03 | — | 📊 leaderboard ↗ |
Operator guidance
Exceptional price-to-performance ratio — one of the cheapest frontier-class models available. Route here when cost is the primary constraint and capability is still required. Available directly via DeepSeek API or on Replicate, Together AI.
Use cases
- Cost-optimised code generation and general reasoning
- RAG and retrieval pipelines where cost is primary constraint
- Open-weights deployment on own infrastructure
Limitations
- Chinese lab; some organisations have data-residency or compliance constraints
- Open weights means self-hosting cost must be factored for on-prem use
- Capability ratings are qualitative, not a single cited benchmark run