Article published in February 2026, updated on 12 November 2026.
In February, we reported rumours of an imminent release of Claude Sonnet 5, with no date and no confirmation whatsoever from Anthropic. The model did indeed come out on 30 June 2026, and it has already been replaced, on 28 September, by Claude Sonnet 5.5, sold at the same price. This update takes stock of what had been assumed, what actually arrived, and what it changes for an SME in French-speaking Switzerland that uses an AI assistant every day. Our position fits in one sentence: the version number matters less than three very down-to-earth choices, namely the contract under which your data is processed, the task you give the model and the real cost of that task.
Claude Sonnet 5: what we wrote in February, and what proved true
The winter rumours revolved around a simple idea: a mid-range model that would come close to the top of the range without its price. On this point, the official announcement proves them right. Anthropic presents Claude Sonnet 5 as “the most agentic Sonnet model to date”, able to plan, use tools such as a browser or a terminal, and work autonomously. The company also writes that its performance is close to that of Claude Opus 4.8, its more general-purpose model, at a lower price.
The price, precisely, came as a surprise. Anthropic launched it at $2 per million input tokens and $10 per million output tokens, whereas Claude Sonnet 4.6 is listed at $3 and $15 in the documentation. This rate was presented as introductory, with a move to $3 and $15 planned for 1 September. A note added to the announcement on 10 August 2026 states that the launch rate has become permanent and that the planned increase no longer applies.
Other promises turned out to be less new than announced. Long context, for example, was not exclusive: Anthropic's documentation already gives Sonnet 4.6 a window of one million tokens. As for speed, the June announcement does not emphasise it. It is Sonnet 5.5, three months later, that Anthropic describes as faster. Finally, for developers, the company raised usage limits in Claude Code as well as in its other products, to absorb the higher consumption of high effort levels.
One detail slipped under the radar of all the rumours, and it matters for the budget. Sonnet 5 uses a new way of splitting text into tokens, so that the same document can represent, according to Anthropic, between 1 and 1.35 times more tokens than before. The price per token has therefore dropped, but the number of tokens billed for the same work may increase. In other words, only a measurement on your own documents tells you whether the bill really goes down.
What Anthropic actually published about Sonnet 5
According to the 30 June announcement, Sonnet 5 became the default model of Claude's free and Pro plans, and it is available on the Max, Team and Enterprise plans. Developers access it through the API under the identifier claude-sonnet-5. The spec sheet adds a context window of one million tokens, around 555’000 words according to Anthropic, and up to 128’000 output tokens.
Two entries in this sheet deserve a business leader's attention. The first is the model's reliable knowledge cutoff, set at January 2026: everything that has happened since, a new law or a change of price at a supplier, is unknown to it unless it finds it in the documents you give it. The second is its status. Sonnet 5 is now classed as “legacy”, which means, in Anthropic's vocabulary, that it no longer receives updates, and its retirement will not take place before 30 June 2027.
On the safety side, Anthropic reports a lower overall rate of undesirable behaviours than Sonnet 4.6, better resistance to instructions hidden in content (prompt injection) and fewer hallucinations. The company does, however, acknowledge in the same announcement a slightly higher rate of misaligned behaviours than Opus 4.8 in its automated audits. These are the company's own evaluations of its own model, and we read them as such.
Since September, Sonnet 5.5 has taken over
The autumn renewed almost the whole range. Anthropic's release notes list Claude Opus 5 on 24 July, Claude Opus 5.5 on 22 September, Claude Sonnet 5.5 on 28 September and Claude Haiku 5.5 on 7 October. At the date of this update, the models overview page highlights four current models:
- Claude Fable 5.2, for demanding reasoning and long-running agent tasks, at $8 and $40 per million tokens;
- Claude Opus 5.5, which the documentation recommends as the starting point for most uses, at $4 and $20;
- Claude Sonnet 5.5, described as the best balance between speed and intelligence, at $2 and $10;
- Claude Haiku 5.5, designed for high-volume tasks such as sorting, extraction or routing, from $0.10 and $0.50.
The same page mentions Claude Mythos models with limited access, reserved for organisations verified through Anthropic's programmes. They are not relevant to an ordinary SME.
Sonnet 5.5 keeps exactly the same price as Sonnet 5. Anthropic presents it as “a clear improvement”, more than 30% faster and up to 30% less costly per task according to its own tests, notably because it generally consumes fewer tokens. The company also stresses the creation of polished documents, presentations and spreadsheets. Its reliable knowledge runs up to June 2026, against January for Sonnet 5. For anyone starting today, we therefore see little reason to choose Sonnet 5 over its successor.
What this changes in practice for an SME in French-speaking Switzerland
Writing and documents
To write a proposal, reword a letter or prepare a presentation, the difference between two generations of Sonnet is felt less than the quality of the instruction and of the documents supplied. Take an accounting firm of ten people in Fribourg that writes year-end cover letters every month. The model can save it time on formatting, but it is unaware of tax changes made after its knowledge cutoff, and it can be wrong with confidence. Review by someone who knows the file therefore remains the rule, whatever the model.
Analysing long documents
The one-million-token window makes it possible to submit a framework agreement and its annexes, or several years of minutes, in one go. This is useful for finding a clause, comparing two versions or drawing up the list of deadlines. That said, a reading capacity is not a guarantee of completeness. We advise asking the model to quote the passages it relies on, so that checking takes a few minutes instead of a full re-read.
High-volume automation
As soon as the task is to sort hundreds of incoming emails, extract the amounts from supplier invoices or classify quote requests, the choice of model becomes a question of cost. The documentation positions Haiku 5.5 for this type of work. It also states that requests processed later in batches benefit from a 50% discount. For an overnight job that does not need an immediate answer, this option often weighs more on the bill than switching from one model to another.
Code and internal tools
The announced gains on code mainly concern teams that use Claude Code or the API to maintain a website, a connector or a small business tool. The point of attention lies elsewhere: applications built in 2025 often call a specific model identifier, and these identifiers have an end date. Anthropic thus announced on 30 September 2026 the retirement of Claude Sonnet 4.5 from its API on 30 November 2026, with Sonnet 5.5 as the recommended replacement. After that date, requests to a retired model fail, which can bring to a sudden halt an automation that no one is monitoring any more.
Data entered: what the nLPD and Anthropic say
The Federal Data Protection and Information Commissioner (FDPIC) reminded us in an update published in May 2025 that Swiss data protection law applies directly to data processing based on AI. In particular, it requires the purpose, the functioning and the data sources to be made transparent, and it demands an impact assessment where the risk is high. The FDPIC adds that a person who converses with a language model has the right to know that they are addressing a machine.
In October 2026, as part of Cyber Security Month, the FDPIC devoted a publication to generative AI at work. The central message is that people often hand over far more information than they think when they enter data into an AI tool, and that in a professional context this is frequently sensitive third-party data: clients, patients, employees. On its page on everyday AI, the FDPIC recommends finding out how the provider uses the data and, where possible, removing personal data before entering a request.
On Anthropic's side, the distinction that matters is the contract. For consumer plans, free, Pro and Max, including Claude Code used from these accounts, a privacy setting determines whether conversations are used to train the models; Anthropic states a retention period of five years if you accept, and of 30 days otherwise. Plans subject to commercial terms, that is Team, Enterprise and the API, are not covered by this setting: according to Anthropic's Privacy Center, the inputs and outputs of these products are not used for training by default. Since July 2026, administrators of an Enterprise plan can also choose which models their users have access to.
Our reading is simple. A team that handles client files should not do so from free personal accounts, whatever the model. The right starting point is a business account, with a short written rule saying what is never pasted into the tool.
Moving to Sonnet 5.5 without nasty surprises
For use in the Claude app, the process is light. Simply check in the selector which model is active, then redo two or three usual tasks to compare the result. For an application or automation connected to the API, a little more method is needed, and we generally proceed in this order:
- identify the models actually being called, using the usage export from the Claude console, which details consumption by key and by model;
- replace any call to a deprecated model as a priority, starting with Sonnet 4.5 before 30 November;
- replay a small set of real cases, about twenty documents for example, on the old and the new model, comparing quality, duration and tokens consumed;
- note the cost per task rather than the price per million tokens, since the way text is split has changed.
This last measurement is also the one that lets you reason in francs. Anthropic bills in dollars, but an SME thinks in time saved and monthly spending, and it is on these two quantities that we prefer to judge a change of model. Our article on calculating the ROI of generative AI details this way of counting.
Ultimately, the Sonnet 5 episode mostly recalls the pace of the sector: a model awaited for months, launched at the end of June, become “legacy” at the end of September. Building a lasting use of AI is less about following every release than about being able to change model without rebuilding everything. This is the approach of our AI and automation support, and we can talk it over with you during a first conversation.
Sources
- Anthropic: announcement of Claude Sonnet 5
- Anthropic: announcement of Claude Sonnet 5.5
- Anthropic: overview of current Claude models
- Anthropic: Claude Sonnet 5 spec sheet
- Anthropic: Claude Sonnet 5.5 spec sheet
- Anthropic: model deprecation and retirement schedule
- Anthropic: release notes for the Claude apps
- Anthropic: update to the consumer terms of use
- Anthropic: use of data for model training
- FDPIC: the current data protection law is directly applicable to AI
- FDPIC: generative AI at work, Cyber Security Month 2026 campaign
- FDPIC: using artificial intelligence in everyday life




