DeepSeek pricing 2026: Bedrock EU vs self-host

TL;DR

DeepSeek pricing for EU processing in 2026 is Bedrock in Stockholm: 0.74 USD input and 2.22 USD output per million tokens, model deepseek.v3.2, in-region only. Google MaaS is global and already deprecated. Foundry lists DeepSeek without readable dollar rates. Self-host only when prompts must stay off the cloud or GPU load stays high.

What does DeepSeek cost on Amazon Bedrock in the EU?

DeepSeek pricing on Amazon Bedrock is on-demand per token with no spend floor. In Europe (Stockholm), DeepSeek V3.2 costs 0.74 USD input and 2.22 USD output per million tokens. US East and US West list 0.62 and 1.85 USD. There is no geo profile. For processing in an EU member state, Stockholm is the documented option.

The Bedrock pricing page puts DeepSeek V3.2 on the standard tier. Stockholm shares the higher table with Mumbai, São Paulo, Jakarta and Tokyo. Sydney has its own rates, including V3.1. We did not read a London price on that page, even though the model card lists eu-west-2. London is also not an EU member state.

Endpoint	Input / 1M	Output / 1M	Residency
Bedrock US East / US West	0.62 USD	1.85 USD	United States
Bedrock Europe (Stockholm), eu-north-1	0.74 USD	2.22 USD	Sweden (EU)
Bedrock eu-west-2 (London)	not read	not read	United Kingdom
Bedrock geo or global profile	not offered	not offered	n/a

The card gives launch on 1 December 2025, EOL no sooner than 1 December 2026, 164K context, 8K max output and a March 2025 knowledge cutoff. Geo and global inference IDs are “Not supported”. You call deepseek.v3.2 in the source Region. Frankfurt is absent. How the platform works is in our Amazon Bedrock EU guide.

We recommend Stockholm when prompts should stay in the Union. The premium versus US rates is about 19 percent on input and 20 percent on output. That is a regional list, not Claude’s flat 10 percent add-on.

DeepSeek vs Claude: which token price applies in 2026?

DeepSeek V3.2 in Stockholm is cheaper per million tokens than Claude Sonnet 5.5. Anthropic lists Sonnet 5.5 at 2 USD input and 10 USD output. Regional and multi-region endpoints from Sonnet 4.5 add 10 percent, so 2.20 / 11 USD. The same million output tokens cost 2.22 USD on DeepSeek Stockholm and 11 USD on Sonnet 5.5 in the EU.

The Anthropic pricing page is the Claude API list and the base for cloud partners. Opus 5.5 is 4 / 20 USD. Haiku 5.5 starts at 0.10 / 0.50 USD for prompts up to 100,000 tokens. The full Claude table is in Claude API pricing in the EU.

Model and endpoint	Input / 1M	Output / 1M	EU processing
DeepSeek V3.2, Bedrock Stockholm	0.74	2.22	Yes, in-region
DeepSeek V3.2, Bedrock US	0.62	1.85	No
Claude Sonnet 5.5, global	2.00	10.00	No
Claude Sonnet 5.5, regional / multi-region	2.20	11.00	Yes, on Bedrock EU or Google EU
Claude Opus 5.5, global	4.00	20.00	No

Worked example, our assumption, not a vendor quote: 10 million input and 2 million output tokens per month.

Route	Maths	USD / month
DeepSeek V3.2 Stockholm	10 × 0.74 + 2 × 2.22	11.84
DeepSeek V3.2 US	10 × 0.62 + 2 × 1.85	9.90
Sonnet 5.5 global	10 × 2.00 + 2 × 10.00	40.00
Sonnet 5.5 regional	10 × 2.20 + 2 × 11.00	44.00

DeepSeek is the cheaper token. Keep Claude when you need the model, the tool stack or Anthropic’s processor terms. The contract path across Bedrock, Google and Foundry is in our Claude GDPR comparison.

What do Google Cloud and Microsoft Foundry list?

Google lists DeepSeek V3.2 as a managed API, but only on the global endpoint and already on a deprecation clock. Microsoft Foundry catalogues DeepSeek V3.2 and V4 variants. The public pricing page hid the dollar amounts. Neither route replaces Stockholm on Bedrock for EU residency.

The MaaS page warns that every DeepSeek MaaS endpoint is deprecated or retired. deepseek-v3.2-maas has been deprecated since 21 July 2026 and retires on 21 October 2026. The V3.2 card lists only global, 163,840 context and 65,536 max output. V3.1, R1 and OCR sit in us-central1. We did not read an EU multi-region price there.

Foundry sells DeepSeek directly. The Azure model page lists DeepSeek-V3.2 and DeepSeek-V3.2-Speciale at 128,000 input and output tokens, plus DeepSeek-V4-Pro and DeepSeek-V4-Flash at 1,000,000 input and 384,000 output. The Foundry pricing page showed placeholders on 12 October 2026. We therefore quote no Foundry dollar rates.

If you front several providers with one gateway, use LiteLLM as an EU gateway and pin the region.

When does self-hosting beat Bedrock?

Self-hosting wins when prompts must not leave your network or when load covers GPU fixed costs every month. The open weights are MIT licensed. Bedrock V3.2 is the cheaper start because you buy no cards. V4.1 Flash is not on Bedrock. You run it locally or on another cloud.

The DeepSeek-V3.2 card lists 685B parameters and MIT. DeepSeek-V4.1-Flash has a 552B backbone, 8B active in prefill and 16B in decode, plus 196B Engram, a 1 million token context and 890 bytes of global KV cache per token, also MIT. Those cards do not publish GPU hourly rates. Server sizing is in DeepSeek V4.1 Flash self-hosted.

Path	What you pay	Residency	Model
Bedrock Stockholm	0.74 / 2.22 USD per 1M tokens	In-region Sweden	V3.2, 164K / 8K
Google MaaS	price not read, endpoint deprecated	global only	V3.2 until 21 October 2026
Foundry	dollar amounts not readable	no EU zone shown	V3.2 and V4 in the catalogue
Own GPUs	power, depreciation, operations	your data centre	V3.2 or V4.1 Flash

We recommend Bedrock Stockholm for the first production DeepSeek service in the EU. Self-host V4.1 Flash only when in-region is not enough for legal or you need the newer checkpoint.

How do you call DeepSeek V3.2 in Stockholm?

You need an AWS account, Bedrock access in eu-north-1 and model access for deepseek.v3.2. Then you call the Converse API in that Region. There is no eu. geo prefix for DeepSeek. Five steps cover the first run if IAM and SCPs already allow Stockholm.

1. In the Bedrock console, select Europe (Stockholm).
2. Open Model access and enable DeepSeek V3.2.
3. Create an IAM role or Bedrock API key with bedrock:InvokeModel on eu-north-1.
4. Set AWS_DEFAULT_REGION=eu-north-1. Do not call from Frankfurt and expect failover.
5. Send the first Converse call with the bare ID deepseek.v3.2.

import boto3

client = boto3.client("bedrock-runtime", region_name="eu-north-1")
response = client.converse(
    modelId="deepseek.v3.2",
    messages=[
        {
            "role": "user",
            "content": [
                {"text": "List five duties of a GDPR processor in short bullets."}
            ],
        }
    ],
)
print(response["output"]["message"]["content"][0]["text"])

The model card uses the same Converse shape with us-east-1 in the sample. Swap the Region. max_tokens above 8,000 is rejected. Point budget alarms at the DeepSeek marketplace line, not only at Amazon Bedrock.

FAQ

How much does DeepSeek cost for an internal chatbot?

Our assumption, 40,000 requests per month, 4,000 input and 500 output tokens each: 160 million input and 20 million output. On Bedrock Stockholm that is 160 × 0.74 + 20 × 2.22 = 162.80 USD. The same load on Sonnet 5.5 regional is 160 × 2.20 + 20 × 11 = 572 USD. That is not a quote.

DeepSeek vs Claude: which model is cheaper for EU teams?

DeepSeek V3.2 in Stockholm is cheaper per token. Sonnet 5.5 regional costs 2.20 / 11 USD, DeepSeek 0.74 / 2.22 USD. Claude wins when you need the Claude tool stack, higher output limits or Anthropic’s contract. Score quality on your tasks, not only on the list price.

Is DeepSeek on Bedrock available in Frankfurt?

No. The model card lists eu-north-1 (Stockholm) and eu-west-2 (London) in Europe. Frankfurt is missing. There is no geo or global profile. If Frankfurt is your only AWS Region, open Stockholm or pick another model.

What does DeepSeek cost on DeepSeek’s own API?

This page only cites AWS, Google Cloud, Microsoft, Anthropic and Hugging Face. We did not read DeepSeek’s first-party price list in this run. For an EU company that needs residency, Bedrock Stockholm is the documented path, not the DeepSeek cloud.

Does Google DeepSeek MaaS run in the EU?

No. deepseek-v3.2-maas is global only. The endpoint has been deprecated since 21 July 2026 and retires on 21 October 2026. V3.1, R1 and OCR run in us-central1. For residency in the Union, use Bedrock Stockholm or a self-hosted checkpoint.

Sources

1. Amazon Bedrock pricing (12 October 2026)
2. AWS: DeepSeek V3.2 model card (12 October 2026)
3. Hugging Face: DeepSeek-V3.2 (12 October 2026)
4. Hugging Face: DeepSeek-V4.1-Flash (12 October 2026)
5. Google Cloud: DeepSeek models (MaaS) (12 October 2026)
6. Google Cloud: DeepSeek-V3.2 (12 October 2026)
7. Google Cloud: Open model deprecations (12 October 2026)
8. Anthropic: Pricing (12 October 2026)
9. Microsoft Learn: Foundry models sold by Azure (12 October 2026)