R1 vs DeepSeek V3.2
R1 and V3.2 are both DeepSeek models but built almost a year apart (January 2025 vs December 2025), and it shows in pricing and design. R1 forces reasoning on every call and caps output at 16,000 tokens. V3.2 makes reasoning optional, adds sparse attention for efficiency, supports prompt caching, and costs a fraction of R1's price. This page breaks down the actual cost and capability gap so you know which one to route traffic to.
Spec vs spec
| Spec | R1 | DeepSeek V3.2 |
|---|---|---|
| Context window | 164K | 131K |
| Max output | 16K | 64K |
| Input modalities | text | text |
| Output modalities | text | text |
| Knowledge cutoff | Jul 31, 2024 | — |
| Released | Jan 20, 2025 | Dec 1, 2025 |
| Reasoning | always on | optional |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
deepseek/deepseek-r1
Input · 1M tokens
$0.700 + 3%$0.721
Output · 1M tokens
$2.50 + 3%$2.58
FEE — FLAT, EVERY MODEL3%
deepseek/deepseek-v3.2
Input · 1M tokens
$0.229 + 3%$0.236
Output · 1M tokens
$0.343 + 3%$0.353
Cache read · 1M tokens
$0.023 + 3%$0.024
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
deepseek/deepseek-r1
$12.36
$12.00 provider + 3%
deepseek/deepseek-v3.2Cheaper
$3.06
$2.97 provider + 3%
Pricing math
R1 costs $0.70/M input and $2.50/M output from the provider; through OpenKey that's $0.721/M and $2.575/M after the flat 3% fee ($0.70 x 1.03, $2.50 x 1.03). V3.2 costs $0.2288/M input and $0.3432/M output provider-side, or $0.235664/M and $0.353496/M on OpenKey (same 3% math). On a 10M-input/2M-output workload, R1 runs $12.00 while V3.2 runs $2.97 — R1 is roughly 4x the cost for identical volume. V3.2 also supports cache reads at $0.02288/M, something R1 doesn't offer at all, which cuts repeat-context costs further on multi-turn workloads.
Reasoning behavior
R1 has reasoning marked mandatory — every request runs the full chain-of-thought, whether you want it or not. V3.2 makes reasoning optional and disabled by default, so you choose per-request whether to pay for the extra reasoning tokens. If your app needs fast, cheap completions most of the time and deep reasoning occasionally, V3.2's toggle is the better fit. If you specifically want R1's open reasoning traces for research or debugging model thought process, that's a reason to stick with R1.
Context and output limits
R1 has a larger context window at 163,840 tokens versus V3.2's 131,072 — a 1.25x ratio in R1's favor. But V3.2 flips that advantage on output: its max completion is 64,000 tokens against R1's 16,000, a 4x difference. For long-document ingestion where the window matters more than the reply length, R1 has a slight edge. For workloads that generate long outputs — long-form code, reports, multi-step agent plans — V3.2's completion ceiling matters more.
Capability signals
Artificial Analysis benchmarks put R1 at an intelligence index of 18.5, a coding index of 24.6, and an agentic index of 3.1. V3.2 doesn't have artificial_analysis scores in this dataset, but it does have Design Arena rankings across eight categories — for example rank 41 in 3D (elo 1210, 50% win rate) and rank 46 in website (elo 1217, 50.6% win rate), with its weakest category being SVG at rank 54 (elo 1089, 41% win rate). These aren't directly comparable metrics, so treat capability claims here as separate data points rather than a head-to-head score.
When to pick each
Pick V3.2 for anything cost-sensitive at scale, anything needing long outputs, or anything where you want reasoning to be optional rather than forced. Pick R1 when you specifically need its mandatory open reasoning tokens, when your prompts lean on the larger 163,840-token context, or when you're benchmarking against its known intelligence/coding/agentic index scores. Both models run on OpenKey with one API key and the same flat 3% fee on top of provider pricing, so switching between them is a model-ID change, not a re-integration.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| High-volume API traffic | DeepSeek V3.2 | Costs $2.97 vs $12.00 for a 10M-input/2M-output workload |
| Long-form code or report generation | DeepSeek V3.2 | 64,000 max output tokens vs R1's 16,000 |
| Forced chain-of-thought reasoning traces | R1 | Reasoning is mandatory on every call, unlike V3.2's optional toggle |
| Very long input documents | R1 | 163,840 token context vs V3.2's 131,072, a 1.25x edge |
| Multi-turn apps with repeated context | DeepSeek V3.2 | Supports cache reads at $0.02288/M; R1 has no caching |
| Matching against known intelligence/coding benchmarks | R1 | Has published Artificial Analysis scores (18.5 intelligence, 24.6 coding, 3.1 agentic) |
Questions
- Which is cheaper, DeepSeek R1 or V3.2?
- V3.2 by a wide margin. Input runs $0.2288/M vs R1's $0.70/M, and output runs $0.3432/M vs R1's $2.50/M. On a 10M-input/2M-output workload that's $2.97 for V3.2 versus $12.00 for R1 — roughly 4x cheaper.
- Does R1 or V3.2 have a bigger context window?
- R1 does, at 163,840 tokens versus V3.2's 131,072 tokens, a 1.25x ratio. But V3.2 wins on max output: 64,000 tokens versus R1's 16,000, so it can generate much longer single responses.
- Is reasoning mandatory on both models?
- No. R1 has reasoning marked mandatory on every request. V3.2 makes reasoning optional and disabled by default, so you pay for reasoning tokens only when you explicitly enable it.
- Can I use both models through the same API key?
- Yes. Both R1 and V3.2 run on OpenKey with a single API key, and pricing is the provider's list price plus a flat 3% fee — for example R1's $0.70/M input becomes $0.721/M on OpenKey.