Gemma vs Claude: EU pricing, Frankfurt 2026

TL;DR

Gemma vs Claude in the EU is residency first. Claude Sonnet 5.5 costs 2 / 10 USD per million tokens plus 10 percent on regional endpoints and stays behind eu.anthropic.claude-sonnet-5-5. Gemma 4 31B in Frankfurt is 0.17 / 0.48 USD, has no geo profile and runs only on bedrock-mantle.

What separates Gemma and Claude for a company?

Gemma vs Claude in 2026 is two product lines with opposite residency models. Claude Sonnet 5.5 gives 1 million tokens of context behind an EU geo profile on Bedrock. Gemma 4 gives Apache 2.0 weights and stays in the single Region you call. None of the Gemma 4 cards we read publish a geo ID. Frankfurt is the documented EU row.

The Sonnet 5.5 card dated 28 September 2026 lists a 1 million token context, 128,000 output tokens and a June 2026 cutoff. The EU profile is eu.anthropic.claude-sonnet-5-5. Model EOL is no sooner than 28 September 2027. Which Claude model to start with is in Claude Sonnet vs Opus.

The Gemma 4 31B card lists 30.7 billion parameters, launch on 31 March 2026, EOL no sooner than 31 March 2027, 256,000 context tokens and runtime ID google.gemma-4-31b. Geo and global are “Not supported”. Gemma 4 runs only on bedrock-mantle under /openai/v1, not /v1. The 26B-A4B card lists 25.2 billion parameters total, 3.8 billion active and ID google.gemma-4-26b-a4b. The E2B card lists 5.1 billion total, 2.3 billion effective, 128,000 context tokens and ID google.gemma-4-e2b.

Trait	Claude Sonnet 5.5	Gemma 4 31B	Gemma 4 26B A4B	Gemma 4 E2B	Gemma 3 27B
Role	Production, agents	Dense open-weight	Fast MoE	Compact	Legacy open-weight
Launch	28 September 2026	31 March 2026	31 March 2026	31 March 2026	12 March 2025
Context / output	1M / 128K	256K / not read	256K / not read	128K / not read	128K / 8K
Bedrock ID	eu.anthropic.claude-sonnet-5-5	google.gemma-4-31b	google.gemma-4-26b-a4b	google.gemma-4-e2b	google.gemma-3-27b-it
EU on Bedrock	Geo profile, active	In-region Frankfurt	In-region Frankfurt	In-region Frankfurt	In-region, no geo
Weights	Closed	Apache 2.0	Apache 2.0	Apache 2.0	Gemma licence (3 family)
Lifecycle	Active	Active	Active	Active	Legacy, EOL 30 March 2027

The Gemma 3 27B card lists launch on 12 March 2025, model EOL 30 March 2027, lifecycle Legacy, 128,000 context tokens, 8,000 output tokens and ID google.gemma-3-27b-it. The 12B card and the 4B card share that launch, EOL and Legacy status. The runtime IDs are google.gemma-3-12b-it and google.gemma-3-4b-it. “EOL no sooner than” 12 March 2026 is a floor that has already passed, not a published shutdown date. The hard date is 30 March 2027.

The Hugging Face Gemma 4 31B-it card lists Apache 2.0, 30.7 billion parameters, 60 layers, 256K context and a 262K vocabulary. 26B A4B has 8 active and 128 experts plus one shared expert. The Gemma 3 27B card lists 14 trillion training tokens, 128K input, 8192 output, more than 140 languages and images at 896 by 896 encoded as 256 tokens. Transformers from 4.50.0. The other Apache 2.0 open-weight comparisons are Qwen vs Claude and Mistral vs Claude. Llama 4 stays on a Community License, see Llama vs Claude.

What does Gemma vs Claude cost in the EU?

Gemma vs Claude on Bedrock is an order-of-magnitude price gap. Anthropic lists Sonnet 5.5 at 2 / 10 USD per million tokens, regional endpoints at 10 percent more, so 2.20 / 11 USD in the EU. Gemma 4 31B in Frankfurt is 0.17 / 0.48 USD. That is the documented EU rate for the dense 31B.

On-demand standard prices on 16 October 2026, USD per million tokens:

Route	Input / output	Cache read	Note
Claude Sonnet 5.5, Claude API global	2.00 / 10.00	0.10	5-minute write 2.50, 1-hour write 4.00
Claude Sonnet 5.5, Bedrock EU or Google Europe	2.20 / 11.00	0.11 (10 percent rule)	Batch is 50 percent of the EU row
Claude Opus 5.5, Claude API global	4.00 / 20.00	0.20	regional endpoints plus 10 percent
Claude Haiku 5.5, prompts up to 100,000 tokens	0.10 / 0.50	0.01	above that 0.50 / 2.50
Gemma 4 31B, US East/Ohio/West	0.14 / 0.40	not read	no geo profile
Gemma 4 31B, Frankfurt	0.17 / 0.48	not read	only Gemma 4 EU row
Gemma 4 26B A4B, Frankfurt	0.16 / 0.48	not read	3.8B active
Gemma 4 E2B, Frankfurt	0.05 / 0.10	not read	2.3B effective
Gemma 3 27B, Ireland and Milan	0.27 / 0.45	not read	Legacy
Gemma 3 27B, London	0.36 / 0.59	not read	highest Gemma 3 EU row
Gemma 3 12B, Ireland and Milan	0.11 / 0.34	not read	Legacy
Gemma 3 4B, Ireland and Milan	0.05 / 0.09	not read	Legacy

The Bedrock pricing page sets Priority at a 75 percent premium and Flex and Batch at a 50 percent discount to Standard. For Gemma 3 this run showed no Frankfurt dollar table. We do not invent a Frankfurt rate for 27B, 12B or 4B. The full Claude matrix is in Claude API pricing in the EU.

For 10,000 support tickets at 3,000 input and 400 output, our arithmetic on the published rates: Claude Sonnet 5.5 EU 110 USD, Gemma 4 31B Frankfurt 7.02 USD, Gemma 4 26B Frankfurt 6.72 USD, Haiku 5.5 (prompts under 100,000 tokens) 5 USD. The price gap is real. So is the ops gap: Gemma 4 speaks a different API.

Does inference stay in the EU?

Gemma vs Claude stays in the EU only if you pin the right ID and the right host. Claude Sonnet 5.5 stays behind eu.anthropic.claude-sonnet-5-5. Gemma 4 has no eu.google. profile. The residency control is the Region you set on the client, plus the Mantle host bedrock-mantle.eu-central-1.api.aws.

From Frankfurt the Claude EU profile routes to Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. London as a source can also hit London. That is Europe, not a pin on one data centre. How Bedrock geo profiles work is in the Amazon Bedrock EU guide.

The compatibility page lists us-east-1, us-east-2, us-west-2, us-gov-west-1 and eu-central-1 for Gemma 4 31B, 26B-A4B and E2B. That is the only EU Region on that list. Gemma 3 27B, 12B and 4B list eu-south-1, eu-west-1 and eu-west-2 among others. The model cards also show Frankfurt and Stockholm in the glance table. The pricing page did not ship a Frankfurt dollar rate for Gemma 3.

Google Cloud lists Claude Sonnet 5.5 in the Europe multi-region at 1,250 requests per minute, 12,500,000 uncached input TPM and 1,250,000 output TPM, context 1,000,000. Gemma 4 26B A4B IT MaaS is global only, ML processing is us, context 262,144 tokens, output 128,000 tokens, launch stage Experimental, release 3 April 2026. There is no Europe multi-region on that card.

The Foundry catalogue of models sold by Azure showed no Gemma row this run. Claude on Foundry without an EU data zone is in the Foundry Europe guide and the GDPR comparison.

When do you take Gemma, when Claude?

Take Claude Sonnet 5.5 on Bedrock EU when prompts must stay in member states and you need agents, computer use, Guardrails or 1 million tokens of context. Take Gemma 4 31B in Frankfurt when the dollar rate matters, the weights are Apache 2.0 and your stack speaks Mantle /openai/v1. Do not take Gemma 3 as a new default: Legacy, model EOL 30 March 2027.

1. Write residency into the brief: one member state, EU geography or global.
2. Measure the task on a fixed eval set, not a demo chat.
3. Pinning: Claude only via eu.anthropic.claude-sonnet-5-5 or Google region="eu". Gemma 4 only via bedrock-mantle.eu-central-1.api.aws/openai/v1 and google.gemma-4-31b.
4. Put a spend cap on a LiteLLM gateway so apps do not silently fall back to global. or us-east-1.
5. Plan Gemma 4 with EOL no sooner than 31 March 2027. Plan Gemma 3 with model EOL 30 March 2027.
6. If the weights must sit in your data centre, host Gemma 4 yourself. The 31B-it card lists Apache 2.0 and google/gemma-4-31B-it.

We recommend Bedrock EU for Claude as the default in European companies. Gemma 4 is the price and weight winner in Frankfurt, not the API and agent winner. Gemma 4 does not support parallel tool calls in this run. Reasoning tokens come back only on the Responses API, not Chat Completions. The request body including images and video must stay under 3.5 MB.

How do you call both models on Bedrock?

You need two clients, because Gemma 4 speaks only Mantle /openai/v1 and Claude runs on bedrock-runtime with the EU profile. Set OPENAI_API_KEY to the Bedrock long-term key and AWS_BEARER_TOKEN_BEDROCK or IAM for Converse. No Anthropic key.

import os
from openai import OpenAI
import boto3

gemma = OpenAI(
    api_key=os.environ["OPENAI_API_KEY"],
    base_url="https://bedrock-mantle.eu-central-1.api.aws/openai/v1",
)
g = gemma.chat.completions.create(
    model="google.gemma-4-31b",
    messages=[{"role": "user", "content": "Summarise the ticket in four bullets."}],
)

claude = boto3.client("bedrock-runtime", region_name="eu-central-1")
c = claude.converse(
    modelId="eu.anthropic.claude-sonnet-5-5",
    messages=[{"role": "user", "content": [{"text": "Summarise the ticket in four bullets."}]}],
    inferenceConfig={"maxTokens": 512},
)
print(g.model, c["usage"])

1. Point the Gemma client at eu-central-1 and /openai/v1, not /v1.
2. Pass google.gemma-4-31b or google.gemma-4-26b-a4b, not an eu. ID. There is none.
3. Create the Claude client in eu-central-1 and pass eu.anthropic.claude-sonnet-5-5, not global.anthropic.claude-sonnet-5-5.
4. Log model ID, host and Region next to the request ID, or FinOps only sees a Bedrock account.
5. Gemma 3 can use bedrock-runtime and Converse. Gemma 4 cannot.

Self-hosting follows the Hugging Face card: pip install -U transformers torch accelerate, model ID google/gemma-4-31B-it. We did not read EU rental rates for the GPUs that job needs, so we do not invent them.

FAQ

What does Gemma vs Claude cost for 10,000 support tickets?

On our assumptions (3,000 input, 400 output) Claude Sonnet 5.5 costs 100 USD globally and 110 USD in the EU. Haiku 5.5 up to 100,000 tokens costs 5 USD. Gemma 4 31B in Frankfurt costs 7.02 USD, 26B A4B 6.72 USD. That is our arithmetic on the Bedrock pricing rates from 16 October 2026.

Gemma vs Claude: which model should a European company take first?

Claude Sonnet 5.5 via eu.anthropic.claude-sonnet-5-5 when personal prompts must stay in the EU and you need agents or computer use. Gemma 4 31B in Frankfurt first when token price matters, you want Apache 2.0 weights and the stack speaks Mantle /openai/v1. Gemma 3 is Legacy.

Does Gemma 4 stay in the EU on Amazon Bedrock?

In Frankfurt, if you pin the Region. The model cards from 16 October 2026 publish no geo profile. The compatibility page lists eu-central-1 as the only EU Region for 31B, 26B-A4B and E2B. A call against us-east-1 leaves the EU.

Self-host Gemma 4 or use Google MaaS?

Self-host if the weights must sit in your data centre. The 31B-it card lists Apache 2.0. Google MaaS gemma-4-26b-a4b-it-maas is global only with ML processing us. That is not an EU path.

Claude on Google or Gemma 4 on Google?

Claude Sonnet 5.5 has 1,250 QPM and 1 million tokens of context in the Europe multi-region. Gemma 4 MaaS is global with US processing, 262,144 tokens of context and 128,000 output, status Experimental. There is no matching EU frontier row for Gemma 4 on Google.

Does Gemma 3 27B have a Frankfurt price on Bedrock?

On 16 October 2026 the pricing page showed Gemma 3 dollar tables for Ireland/Milan (0.27 / 0.45), London (0.36 / 0.59) and other non-EU Regions. It did not ship a Frankfurt row for 27B, 12B or 4B. We leave that dollar rate blank.

Sources

1. Anthropic: Pricing (16 October 2026)
2. AWS: Claude Sonnet 5.5 model card (16 October 2026)
3. AWS: Gemma 4 31B model card (16 October 2026)
4. AWS: Gemma 4 26B-A4B model card (16 October 2026)
5. AWS: Gemma 4 E2B model card (16 October 2026)
6. AWS: Gemma 3 27B PT model card (16 October 2026)
7. AWS: Gemma 3 12B IT model card (16 October 2026)
8. AWS: Gemma 3 4B IT model card (16 October 2026)
9. AWS: Regional availability by models (16 October 2026)
10. Amazon Bedrock Pricing (16 October 2026)
11. Hugging Face: gemma-4-31B-it (16 October 2026)
12. Hugging Face: gemma-3-27b-it (16 October 2026)
13. Google Cloud: Gemma 4 26B A4B IT MaaS (16 October 2026)
14. Google Cloud: Claude Sonnet 5.5 (16 October 2026)
15. Microsoft Learn: Foundry models sold by Azure (16 October 2026)