DeepSeek v3.1
DeepSeek-V3.1 via the Vercel AI Gateway: long-context extension (32K→128K), UE8M0 FP8 scale. Post-trained on DeepSeek-V3.1-Base. Served only by zero-data-retention providers. Does not support web search or image input.
You always get the exact model you pick — we never silently route you to another.
About DeepSeek v3.1
DeepSeek V3.1 collapses two separate model endpoints into one: flip a template switch and the same 671-billion parameter architecture goes from fast chat mode to full chain-of-thought reasoning. That architectural consolidation — combined with pricing roughly 68 times lower than Claude Opus — is what developers talk about most. Only 37 billion parameters activate per inference, which keeps latency and cost in check even as benchmark scores climb: 89% on MMLU-Pro, 68.4% on SWE-bench Verified, and AIME 2025 scores competitive with dedicated reasoning models. A 128K context window lets it work through entire codebases or lengthy research papers without chunking. The honest caveat: users running it in production have reported occasional random text insertions and generation artifacts, and it has no image or audio support whatsoever — text is the only input it accepts. For cost-sensitive teams that need serious coding and math capability without juggling two separate model endpoints, V3.1 is a straightforward choice.
Best for
- Software engineering — strong SWE-bench Verified scores make it reliable for code review, completion, debugging, and multi-file refactors
- Mathematical and quantitative reasoning — AIME 2025 performance makes it well-suited for complex math, data analysis, and scientific problem-solving
- Long-document analysis — 128K context handles full codebases, lengthy contracts, or research papers in a single pass
- Agentic and tool-using workflows — post-training improvements to function calling make it a practical backbone for multi-step agent pipelines
- Budget-conscious production use — near-frontier reasoning quality at a fraction of the cost of proprietary alternatives
Specifications
| Provider | DeepSeek |
|---|---|
| Released | 2025-08-21 |
| Intelligence | High |
| Speed | Fast |
| Context window | 163,840 tokens |
| Knowledge cutoff | July 2025 |
| Input price | $0.25 / 1M tokens |
| Output price | $0.95 / 1M tokens |
| Request cost | 2 base requests |
| Plan tier | Base |
| Features | Function calling supported, Served through the Vercel AI Gateway (deepseek/deepseek-v3.1), Zero data retention and no prompt training |
| Model ID | deepseek-v3p1 |
Frequently asked questions
V3.1 runs in two modes from a single model. Standard (non-thinking) mode behaves like a fast chat model; reasoning (thinking) mode activates chain-of-thought. You switch between them via the deepseek-chat and deepseek-reasoner endpoints rather than deploying two separate models.
128,000 tokens context window. Knowledge cutoff is July 2025.
Via the DeepSeek API, input is $0.555 per million tokens and output is $1.67 per million tokens. Third-party providers on OpenRouter or DeepInfra may differ — check the specific provider for current rates.
No. DeepSeek V3.1 is text-only. It cannot process images, audio, video, or any other non-text modality.
Some production users have noted occasional random text insertions and generation artifacts in outputs. The model also occasionally makes errors on certain complex logical puzzles. These are worth testing against your specific workload before committing.
V3.1's reasoning mode (Intelligence Index 60) slightly outperforms R1 (59) while adding faster response times. The key difference is that V3.1 unifies both fast and reasoning modes in one model, whereas R1 was reasoning-only.