Claude Sonnet vs Opus for Company Workloads 2026

TL;DR

Claude Sonnet vs Opus in 2026 is a cost and latency choice, not a residency choice. Sonnet 5.5 costs 2 / 10 USD per million tokens, Opus 5.5 costs 4 / 20 USD. Both have a Bedrock eu. ID and a Google Europe multi-region route. Default Sonnet 5.5 for volume; promote to Opus 5.5 when evals fail.

What is the difference between Claude Sonnet and Claude Opus?

Claude Sonnet vs Opus is now a 5.5-generation choice. Sonnet 5.5 is the fast production model at 2 / 10 USD. Opus 5.5 is the coding and knowledge-work model at 4 / 20 USD. Both have a 1 million token context window, 128,000 token max output, and a June 2026 knowledge cutoff. Anthropic tells you to start on Opus 5.5.

The models overview published on 2 October 2026 names four current models: Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5. Claude Opus 5 and Claude Sonnet 5 are still available as legacy IDs. Opus 5 still costs 5 / 25 USD, so Opus 5.5 is cheaper than its predecessor. Sonnet 5.5 keeps the same 2 / 10 USD list price as Sonnet 5.

Feature	Claude Sonnet 5.5	Claude Opus 5.5
Role	Speed and intelligence mix	Long-running agentic coding and knowledge work
Released	28 September 2026	22 September 2026
Retirement (Anthropic)	Not sooner than 28 September 2027	Not sooner than 22 September 2027
List price (input / output)	2 / 10 USD per MTok	4 / 20 USD per MTok
Context / max output	1M / 128K	1M / 128K
Thinking	Adaptive; between_tools turns off up-front thinking	Adaptive, always on
Default effort	high	medium
Comparative latency	Fast	Moderate
Claude API ID	claude-sonnet-5-5	claude-opus-5-5

One million tokens is about 555,000 words on the tokenizer introduced with Claude Opus 4.7. That tokenizer produces about 30% more tokens than the pre-4.7 tokenizer. A budget copied from a Sonnet 4.x bill will undershoot. Full token tables sit in our Claude API pricing EU comparison.

Two API details change how you migrate. On Sonnet 5.5, setting temperature, top_p or top_k to a non-default value returns a 400 error, and between_tools is the lowest thinking setting. On Opus 5.5, thinking cannot be turned off; forced tool use (tool_choice of any or a named tool) returns an error. Both models add 286 tool-use system prompt tokens when tool_choice is auto.

How much does each model cost per task?

Opus 5.5 costs twice the Sonnet 5.5 list price on both input and output. A 10,000-ticket support queue at 3,000 input and 400 output tokens per ticket costs 100 USD on Sonnet 5.5 and 200 USD on Opus 5.5 globally. The Google Europe multi-region adds 10%: 110 USD versus 220 USD. Batch halves both sides.

Prices below were read on 2 October 2026 from Anthropic pricing and Google Cloud Agent Platform pricing. No vendor publishes euro list prices. Anthropic applies a 10% premium to Bedrock regional endpoints and Google multi-region endpoints from Sonnet 4.5, Haiku 4.5 and Opus 4.5 onward.

Route	Sonnet 5.5 in / out	Opus 5.5 in / out	Cache read	Batch in / out (Sonnet / Opus)
Claude API (global)	2.00 / 10.00	4.00 / 20.00	0.20 / 0.20	1.00 / 5.00 and 2.00 / 10.00
Google Cloud global	2.00 / 10.00	4.00 / 20.00	0.20 / 0.20	see Google table
Google Cloud Europe multi-region	2.20 / 11.00	4.40 / 22.00	0.22 / 0.22	1.10 / 5.50 and 2.20 / 11.00
Bedrock EU geo (10% rule)	2.20 / 11.00	4.40 / 22.00	0.22 / 0.22	50% of the EU row

USD per million tokens. Opus 5.5 cache reads use a 0.05x multiplier (0.20 USD globally). Sonnet 5.5 cache reads use the usual 0.1x (also 0.20 USD, because the input price is half). A 5-minute cache write is 2.50 USD on Sonnet 5.5 and 5.00 USD on Opus 5.5. A 1-hour write is 4.00 and 8.00 USD.

The ticket and review rows below are our arithmetic, not a vendor quote. Token counts already assume the current tokenizer.

Workload (our assumptions)	Tokens per job	Jobs / month	Sonnet 5.5 global	Opus 5.5 global	Same jobs, Google EU
Support reply	3,000 in, 400 out	10,000	100 USD	200 USD	110 / 220 USD
Code review	40,000 in, 2,000 out	1,000	100 USD	200 USD	110 / 220 USD

If the same review job finishes in one Opus 5.5 call but needs two Sonnet 5.5 calls, Opus wins on cost. Measure that on your eval pack before you lock a default. Fast mode on Opus 5.5 is 8 / 40 USD and exists only on the first-party Claude API, not on Bedrock, Google Cloud or Foundry.

Foundry and Claude Platform on AWS bill in Claude Consumption Units at 0.01 USD per CCU. One hundred CCU equals 1.00 USD of Anthropic list fees. That is a billing unit, not a discount.

Where can you run Sonnet and Opus with EU processing?

Both models have a documented EU processing path on Amazon Bedrock and on Google Cloud. Bedrock publishes eu.anthropic.claude-sonnet-5-5 and eu.anthropic.claude-opus-5-5 and states that those IDs keep data within EU regions. Google lists model availability and ML processing in the Europe multi-region for both. Microsoft Foundry has no EU data zone for Claude.

We read the three platform pages on 2 October 2026.

Platform	Sonnet 5.5 ID	Opus 5.5 ID	EU processing
Claude API	claude-sonnet-5-5	claude-opus-5-5	No. First-party inference_geo is global or us
Amazon Bedrock	eu.anthropic.claude-sonnet-5-5	eu.anthropic.claude-opus-5-5	Yes, EU geo. In-region on bedrock-runtime is N/A
Google Cloud	claude-sonnet-5-5	claude-opus-5-5	Yes, Europe multi-region
Microsoft Foundry	claude-sonnet-5-5	claude-opus-5-5	No EU data zone. Hosted on Azure offers Global Standard and US Data Zone

The Sonnet 5.5 Bedrock card lists geo IDs for US and EU plus global.anthropic.claude-sonnet-5-5. The Opus 5.5 card adds AU and JP geos. AWS sample code still calls the global ID from us-east-1. Copy that sample into a Frankfurt account and you leave the EU. IAM, logging and the eu. prefix are in our Claude on AWS Bedrock EU guide.

Google’s Sonnet 5.5 page and Opus 5.5 page both list United States multi-region, Europe multi-region and the global endpoint. Default quota in a multi-region is 1,250 QPM for Sonnet 5.5 and 1,000 QPM for Opus 5.5. The global endpoint doubles those figures (2,500 and 2,000 QPM). Set region="eu" as in our Claude on Vertex AI in Europe guide.

Foundry is the wrong pick if the requirement is EU inference. Microsoft’s hosting comparison places Opus 5.5 and Sonnet 5.5 on Hosted on Azure (Global Standard and US Data Zone) and on Hosted on Anthropic (Global Standard only). There is no Europe data zone for Claude. Anthropic remains the processor on both hosting options. The wider matrix is in our Claude GDPR comparison.

Fable 5.1 is the exception in this family: it has no Bedrock EU geo ID. If a workload needs Fable and EU processing, Google Cloud is the route. That split is documented in our Claude Fable 5.1 Europe guide.

Which model should you default to for company workloads?

Anthropic’s lineup says start with Claude Opus 5.5 and move to Fable 5.1 only when higher effort still falls short. That is the quality default. For a European production API with thousands of similar calls, we default to Sonnet 5.5 and promote where the eval pack fails. Claude Code already starts on Opus 5.5.

Write the rule down. A verbal “use the smart one” becomes an Opus bill within a week.

1. Build a 30 to 50 item eval pack from real tickets, reviews or agent traces. Score quality, not vibes.
2. Run the pack on Sonnet 5.5 at default high effort and on Opus 5.5 at default medium.
3. Keep Sonnet 5.5 where the scores match. Route the failing slice to Opus 5.5.
4. Re-run the pack after any prompt or tool change. A new system prompt can flip the winner.
5. Pin the model ID in code. Do not ship a generic sonnet or opus alias in production.

Workload	Our default	Why
High-volume chat, classification, extraction	Sonnet 5.5	Same 2 / 10 USD as Sonnet 5, half the Opus 5.5 list price
Claude Code and long coding agents	Opus 5.5	Anthropic’s starting model; Claude Code already defaults here
Dense filings, multi-step knowledge work	Opus 5.5 first	Then Fable 5.1 only if evals stall
Overnight batch rewrite	Sonnet 5.5 batch	1 / 5 USD globally, 1.10 / 5.50 USD on Google EU
Need EU processing on AWS	Either, with eu. ID	Both cards publish an EU geo ID, unlike Fable 5.1

Do not stay on Opus 5 to “be safe”. It costs 5 / 25 USD and is now a legacy row. If you already passed evals on Opus 5, re-run them on Opus 5.5 before you renew a capacity plan.

How do you switch models on Bedrock or Google Cloud?

Pin the EU geo ID on Bedrock or the Europe multi-region on Google Cloud, then change one model string. Do not copy the global ID from the AWS sample. From Frankfurt, call eu.anthropic.claude-sonnet-5-5 or eu.anthropic.claude-opus-5-5. On Google Cloud, keep region="eu" and swap claude-sonnet-5-5 for claude-opus-5-5.

import boto3

client = boto3.client("bedrock-runtime", region_name="eu-central-1")

def complete(model_id: str, text: str) -> dict:
    return client.converse(
        modelId=model_id,
        messages=[{"role": "user", "content": [{"text": text}]}],
    )

sonnet = complete("eu.anthropic.claude-sonnet-5-5", "Summarise the ticket in four bullets.")
opus = complete("eu.anthropic.claude-opus-5-5", "Summarise the ticket in four bullets.")
print(sonnet["output"], opus["output"])

1. Create the Bedrock runtime client in an EU region. eu-central-1 is the usual Frankfurt choice.
2. Pass the eu. inference profile, not global.anthropic.claude-sonnet-5-5 and not the bare anthropic.claude-sonnet-5-5 ID on bedrock-runtime.
3. Keep max_tokens explicit. Both models allow 128,000 output tokens on the synchronous API.
4. Log modelId next to request IDs in CloudTrail or your gateway so FinOps can split the 2x price gap.
5. On Google Cloud, install anthropic[vertex], set region="eu", and use the same two Claude API IDs.

A self-hosted LiteLLM gateway can expose one internal model name and map it to Sonnet or Opus behind a spend cap. That is useful when several apps should not each hard-code Bedrock IDs.

FAQ

How much more does Claude Opus 5.5 cost than Claude Sonnet 5.5?

Twice the list price on input and twice on output: 4 / 20 USD versus 2 / 10 USD per million tokens. On Google’s Europe multi-region that is 4.40 / 22 USD versus 2.20 / 11 USD. Batch is half of each pair. Cache reads are 0.20 USD globally on both models, for different reasons: 0.05x of 4 USD on Opus 5.5, 0.1x of 2 USD on Sonnet 5.5.

Claude Sonnet vs Opus: which model should a European company default to?

Default Sonnet 5.5 for high-volume APIs. Default Opus 5.5 for Claude Code and for tasks where your eval pack already fails on Sonnet. Anthropic’s own overview starts most workloads on Opus 5.5. That is the quality advice. The cost advice for a 10,000-call queue is the opposite, unless quality drops.

Do Claude Sonnet 5.5 and Claude Opus 5.5 stay in the EU on Bedrock?

Yes, if you call the EU geo IDs. The Sonnet card publishes eu.anthropic.claude-sonnet-5-5 and says it keeps data within EU regions. The Opus card publishes eu.anthropic.claude-opus-5-5 with the same wording. The global IDs have no residency constraint. In-region on bedrock-runtime is not offered for either 5.5 model.

Is Claude Sonnet 5.5 the same price as Claude Sonnet 5?

Yes on the list price: both are 2 / 10 USD globally and 2.20 / 11 USD on Google EU. The models are not the same API. Sonnet 5.5 rejects non-default temperature and adds between_tools. Migrate with a test, not a find-and-replace of the ID only.

Claude Opus 5 vs Opus 5.5: should we stay on Opus 5?

No, not for a new default. Opus 5.5 is 4 / 20 USD. Opus 5 is still 5 / 25 USD and sits on the legacy list. Re-run evals on 5.5 before you keep paying the older, higher row.

Does Microsoft Foundry keep Sonnet or Opus inference in Europe?

No. Foundry offers Hosted on Azure with Global Standard and a US Data Zone, plus Hosted on Anthropic with Global Standard only. Microsoft does not list an EU data zone for Claude. Use Bedrock EU or Google Europe multi-region if processing must stay in EU member states.

Sources

1. Anthropic docs: Models overview (2 October 2026)
2. Anthropic docs: Claude Opus 5.5 overview (2 October 2026)
3. Anthropic docs: Claude Sonnet 5.5 overview (2 October 2026)
4. Anthropic docs: Pricing (2 October 2026)
5. AWS Bedrock: Claude Opus 5.5 model card (2 October 2026)
6. AWS Bedrock: Claude Sonnet 5.5 model card (2 October 2026)
7. Google Cloud: Claude Opus 5.5 (2 October 2026)
8. Google Cloud: Claude Sonnet 5.5 (2 October 2026)
9. Google Cloud: Agent Platform pricing (2 October 2026)
10. Microsoft Learn: Claude hosting comparison (2 October 2026)