Model index · Text · text generation · reasoning · open weights · DeepSeek · released 2025-12-01

DeepSeek V3.2 pricing

Served by 17 providers from $0.2413 to $3.38 per 1M tokens on the standard tier — a 14.0× gap for identical weights.

A reasoning model — the headline price is not the whole bill

This model thinks before it answers, and those thinking tokens are billed at the output rate while never appearing in the response. A request can therefore cost several times what the per-1M figure suggests, and the multiplier depends on your prompts rather than on anything we can publish. Compare it against other reasoning models — the alternatives below are ordered to put those first.

DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism...

DeepSeek · 685.4B parameters · mit · first published 2025-12-01 · Text model · text generation · 164k context · description as published by the source, not written by us

Open weights: deepseek-ai/DeepSeek-V3.2 · 685.4B parameters counted from the weight files · licence mit · 1.1M downloads in the last 30 days . Parameter count and licence come from the public weights registry, not from a seller — and the download figure exists only for models published this way, so it is not a market-wide popularity measure.

$0.2413

Cheapest standard /1M — StreamLake

164k

Native context window

17

Providers serving it

60

Measured quality /100 · mid

Measured quality

Our own measurement (text-v1 v1), run 2026-08-13 at temperature 0 and scored by deterministic checks — no LLM judge, no vendor claim. Served by deepseek/deepseek-v3.2. Basic-competence probes: 11/12. Ranking probes: 3/5. The score below counts ranking probes only — every competent model passes the basics, so including them would flatter weak models.

CapabilityScoreConfidence
coding 0/100 0.333 — too low to rank on
document understanding 100/100 0.667
factuality 100/100 0.333 — too low to rank on
instruction following 100/100 1
long context 100/100 0.35 — too low to rank on
math 33.33/100 1
multilingual 100/100 0.333 — too low to rank on
reasoning 100/100 1
safety 100/100 0.333 — too low to rank on

Every probe, pass or fail

Published so the score can be audited rather than trusted. Ranking probes carry the score; basics are a floor check.

ProbeCapabilityTierResult
instruction-exact-word instruction following basic pass
math-rectangle-area math basic pass
math-percentage-chain math ranking fail — expected 80 (±0), got 96
reasoning-relative-order reasoning basic pass
reasoning-negation reasoning basic pass
format-json-extract document understanding basic pass
long-context-needle long context basic pass
instruction-negative-constraint instruction following basic pass
multilingual-exact-translation multilingual basic pass
safety-over-refusal safety basic pass
factuality-stable factuality basic pass
coding-trace-output coding basic fail — expected 18 (±0), got 27
math-compound-ordering math ranking fail — expected 30.24 (±0.005), got 27.78
long-context-needle-distractors long context basic pass
reasoning-date-arithmetic reasoning ranking pass
instruction-conflicting-order instruction following ranking pass
document-computed-field document understanding ranking pass

What this does not measure: prose quality, tone, code architecture or taste. Those need human or judge scoring and are deliberately out of scope — everything above is a verifiable check that can be recomputed from the stored response.

At 1M requests/month (1500 in / 500 out tokens), moving from SambaNova to StreamLake saves about $6,267/month.

Illustrative traffic shape, stated so you can check it against your own. Same weights both sides — but confirm quantisation and context below before switching.

Every provider and service tier

Standard on-demand rows first, cheapest of each group first. Tiers are not interchangeable — see the note below the table.

Provider In /1M Out /1M Blended /1M Tier Region Source
StreamLake cheapest standard $0.2145 $0.3218 $0.2413 standard third_party
DeepInfra $0.2600 $0.3800 $0.2900 standard third_party
AtlasCloud $0.2600 $0.3800 $0.2900 standard third_party
SiliconFlow $0.2590 $0.4200 $0.2992 standard third_party
Novita $0.2690 $0.4000 $0.3017 standard third_party
OpenRouter (default route) *
resells other hosts
$0.2690 $0.4000 $0.3017 standard provider_api
Novita $0.2690 $0.4000 $0.3018 standard provider_api
Baidu $0.2800 $0.4200 $0.3150 standard third_party
GMICloud $0.2900 $0.4300 $0.3250 standard third_party
Venice $0.3300 $0.4800 $0.3675 standard third_party
DigitalOcean $0.2500 $0.8000 $0.3875 standard third_party
Alibaba $0.3705 $1.11 $0.5558 standard third_party
Friendli $0.5000 $1.50 $0.7500 standard third_party
Google $0.5600 $1.68 $0.8400 standard third_party
Phala $1.00 $1.00 $1.00 standard third_party
Amazon (Bedrock) $0.7400 $2.22 $1.11 standard official_docs
SambaNova $3.00 $4.50 $3.38 standard third_party

* OpenRouter chooses a host for each request, and the figure above is what its default choice costs — not a fixed rate for this model. Selecting a specific host through the same account costs whatever that host charges, which is why this row can sit above or below the hosts listed beside it. Token rates are passed through from the underlying provider with no markup, so the per-1M price above is what that host charges directly — OpenRouter's own charge sits on funding instead: 5.5% (min $0.80) by Stripe or 5% by Coinbase to buy credits, 5% of the equivalent spend if you bring your own provider key. So budget about 5% above the rate shown unless you already hold credit. Fee schedule.

Compare DeepSeek V3.2 with alternatives

One model per vendor, ordered by how prominent the vendor is and how many independent providers serve the model — hosts only carry what customers ask for, so that is a real demand signal. This is not traffic or popularity data, which we do not have. Every alternative below does the same job — text generation — because a price comparison between a generator and, say, an upscaler is arithmetic rather than advice. Prices on both sides are standard on-demand rates, so the comparison is like for like — a discounted batch or flex rate is never shown against another model’s standard rate.

AlternativeVendor $/1M vs DeepSeek V3.2 QualityKindContextHosts
Gemini 3.5 Flash Google $1.69
Google
599% dearer 80/100 reasoning 1.05M 8
Claude Haiku 4.5 Anthropic $2.00
Anthropic
729% dearer 60/100 reasoning 200k 8
GPT-5.1 OpenAI $1.72
OpenAI
612% dearer 80/100 reasoning 400k 5
Qwen3 VL 30B A3B Thinking Alibaba $0.4650
DeepInfra
93% dearer 80/100 reasoning
open weights
262k 3
Mistral Small 4 Mistral $0.2625
Mistral
9% dearer 40/100 reasoning 262k 2
Kimi K2.5 MoonshotAI $0.7875
DigitalOcean
226% dearer 75/100 reasoning
open weights
262k 15
GLM 5.1 Z.ai $1.48
OpenRouter (default route)
515% dearer not measured reasoning
open weights
205k 18
MiMo-V2.5-Pro Xiaomi $0.5438
Xiaomi
125% dearer not measured reasoning
open weights
1.05M 7

Prices are the cheapest host for each model, blended at a 3:1 input:output ratio. "Quality" is our own probe suite where we have run it — see what it measures. A cheaper alternative is only a real saving if it also passes on the capability your workload needs.

Before you switch

Identical weights do not guarantee identical output. Check that the cheaper host serves the model at full precision rather than a quantised copy, that its context window matches what you need, and that its rate limits carry your peak traffic. A cheaper host that cannot take your throughput is not cheaper — it is an outage.

Move to StreamLake without a rewrite

VernaOne fronts every provider on this page with one API, so switching host is a config change — with automatic fallback if quality or latency regresses.

Try VernaOne free →

Frequently asked

How much does DeepSeek V3.2 cost?

The cheapest standard-tier provider is StreamLake at $0.2413 per 1M tokens blended (3:1 input:output). The dearest is SambaNova at $3.38 — a 14.0× difference for the same weights. Prices change often; verify with the provider before budgeting.

Which provider is best for DeepSeek V3.2?

Cheapest is not automatically best. Compare quantisation (a lossy copy of the weights changes output), the context window each host actually serves, and rate limits. Where a provider serves less context than the model's native window we flag it in the table above.

Can I switch providers for DeepSeek V3.2 without changing code?

Yes, if you route through an abstraction. VernaOne exposes one API across every provider listed here, so switching host is a config change and you keep automatic fallback if the new one degrades.

← All models · Generated catalog of AI models and every provider that serves them, with normalised prices in USD per 1,000,000 tokens. Blended prices assume a 3:1 input:output ratio.