Is Claude Haiku 5.5 safe? What it does with your data

Short answer: Claude Haiku 5.5 is the smallest and cheapest model in Anthropic’s current line-up, and its launch page says nothing about data handling at all. That silence is not a red flag and not a reassurance. Where your prompts go depends on the plan you use it on, not on the model, and that is the same as it was a week ago.

This post sticks to what is written down. It covers what Anthropic’s launch page says, what Amazon says about running it on Bedrock, and which of Anthropic’s existing data rules apply. It does not summarise the model’s system card. That document is long, and a safety verdict that comes from skimming one is worth less than the card itself. Read it if the model will touch anything sensitive.

What was launched, and when

Anthropic’s page for the model is dated 7 October 2026, and the model ID is claude-haiku-5-5. The page says it is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.

The pricing is the headline. For prompts up to 100,000 tokens, which Anthropic says covers about 90% of requests, the price is $0.10 per million input tokens and $0.50 per million output tokens. Above 100,000 tokens it rises to $0.50 and $2.50. All of this is in US dollars, from Anthropic’s page on 8 October 2026. Amazon’s Bedrock announcement puts the saving at around 75 percent against Claude Haiku 4.5 for most tasks.

A model this cheap changes behaviour. People hand it bigger and more repetitive jobs: every row of a spreadsheet, every email in a folder, every support ticket from last month. That is where the data question stops being theoretical. If you are not sure how a model differs from the product wrapped around it, AI agent vs LLM covers it in a few minutes.

What the launch page leaves out

The launch page does not state a retention period, whether customer data is used for training, whether zero data retention is offered for this model, or where processing happens. Those facts exist, but they sit in other documents, and a launch announcement is not the place a vendor usually repeats them. Do not read the gap as a promise either way. It means the question has to be answered somewhere else.

Where the answer actually lives

Data handling follows the account. On Claude Free, Pro and Max, Anthropic’s consumer terms let you choose whether your chats are used to train future models. If you allow it, chats are kept for up to five years. If you do not, retention is 30 days. The same page says those consumer terms do not apply to Claude for Work, Claude for Government, Claude for Education or API use, including through Amazon Bedrock and Google Cloud’s Vertex AI. Those run under commercial terms.

The setting is called Help improve our AI models and sits under Privacy in your account. Anthropic’s privacy centre explains that turning it off stops new chats going into future training. It does not pull data already used out of a model that has been trained, and conversations flagged by safety classifiers can still be used for trust and safety work.

Amazon’s announcement adds one concrete detail for the cloud route. Haiku 5.5 is offered on Amazon Bedrock through US, EU, Australian, Japanese and Global cross-region inference profiles, and Amazon says Bedrock keeps your data within AWS infrastructure with regional data residency. Amazon’s post does not say whether prompts are shared with Anthropic or used for training, so do not infer it. Check the Bedrock data-protection terms for your own account.

If you are in the European Economic Area, the UK or Switzerland, Anthropic’s privacy centre names Anthropic Ireland, Limited as the data controller for consumer accounts. That is who a GDPR access or deletion request goes to.

What we could not confirm

• —Whether Haiku 5.5 is available inside the consumer Claude apps. The launch page names developer platforms and clouds, and says nothing on this.
• —Whether zero data retention applies to this model. Anthropic’s page for the larger Sonnet 5.5 mentions it. This page does not.
• —Anything in the system card. We have pointed to it and have not summarised it.

What to do before you use it at work

1. 01Find out which plan you are on. A personal Pro account and a company Claude for Work account follow different terms with the same model underneath.
2. 02Look at the training toggle once, on purpose, whichever way you set it.
3. 03Keep secrets out of prompts. API keys, passwords and customer lists belong in a password manager or a controlled system, not in a chat. Is Claude Sonnet 5.5 safe explains why that matters even more with agents.
4. 04If you pay per token and process lots of files, ask your IT team which route you are on, direct API or a cloud provider, because the data terms differ.
5. 05Check the output. A cheap, fast model produces confident wrong answers just as quickly as a dear one. How to write a prompt that works first try helps, but it does not replace reading the result.

For how other vendors handled the same questions this month, is GPT-6 Sol safe and is Mistral Large 4 safe read their documentation the same way. The file-upload question comes up in every assistant, and uploading Excel to ChatGPT answers it for one of them. If you are choosing between assistants rather than models, Muse Spark vs Claude compares plans side by side.

Coursium is a mobile app that teaches people to use AI at work, and it is on the App Store. Model names change every few weeks. Checking where your data goes, keeping secrets out of prompts and reviewing the output carry over from one model to the next. If that is what you want to practise, have a look at Coursium.