Claude Sonnet 5.5: which model for which work, and should you migrate

What Anthropic announced on September 28

Six days after Claude Opus 5.5, Anthropic released Claude Sonnet 5.5, the second model in the 5.5 family. The company presents it as a clear upgrade over Sonnet 5: more than 30% faster, and up to 30% cheaper for most work. Claude Haiku 5.5, built for high-volume and cost-sensitive uses, is due to join the family in the coming weeks.

The model is available now on the Claude Platform and at the three major cloud providers: Amazon Web Services, Google Cloud and Microsoft Azure. Like Opus 5.5 and Sonnet 5, it is offered with a zero data retention option.

For a company already running agents or automations on Sonnet 5, the question is not whether Sonnet 5.5 is better. According to Anthropic, it is, on almost every published test. The real question is practical: should you migrate now, how do you check it on your own cases, and how do you split the work between Sonnet and Opus. That is what this article covers, based on Anthropic's announcement, read in full.

Sonnet or Opus: the split Anthropic proposes

Anthropic describes each model's place clearly. Opus 5.5 is built for complex work that requires careful judgment. Sonnet 5.5 is its faster, lower-cost complement, strongest on well-scoped everyday tasks, bug fixing, and producing polished documents, slides and spreadsheets.

Anthropic stresses a nuance that rankings hide. On several tests, Sonnet 5.5 at maximum effort comes close to Opus 5.5. But in its own trials as in those of external testers, Opus 5.5 remains clearly stronger at open-ended, complex work that requires sustained judgment.

In business terms, this gives a simple rule. A task with a clear scope, where you can say what a good result looks like, and that repeats: that is Sonnet's ground. A task that requires trade-offs, interpreting an ambiguous context or holding long reasoning without precise instructions: that is still Opus's ground. The rule applies to choosing a model as much as to designing an agent, which we detail in deploying an AI agent in your company, step by step.

What the numbers say, and what they do not

The published results are striking on some tests. On GDPval-AA, which evaluates real tasks across 44 occupations, Sonnet 5.5 scores 1,844 points, against 1,449 for Sonnet 5 and 1,846 for Opus 5.5. On OSWorld 2.1, which measures computer use, it reaches 80.1%, against 57.0% for Sonnet 5 and 81.8% for Opus 5.5. On coding tests, the gap with Sonnet 5 is also very wide.

Three precautions apply before making a decision.

• These are the vendor's figures. Some tests are run by third parties, but the selection and presentation are Anthropic's. The announcement's footnotes also flag caveats, such as a bug in the pre-release deployment or competitor scores not yet updated.
• Effort level changes everything. The best scores are often obtained at the highest settings, which cost more and take longer. A ranking score is not the result you will get at the setting you choose.
• Your cases are not the test's. A model that excels on a coding test can get your contracts or your customer requests wrong. Only a test set built on your data tells you what the model is worth to you.

Anthropic says so itself: benchmark scores capture only one facet of a model's capabilities. We developed this point when Opus was released, in Claude Opus 5.5: what it changes for your agents in production.

Cost: same price, fewer tokens

Sonnet 5.5 keeps Sonnet 5's price: 2 dollars per million input tokens, 10 dollars per million output tokens, 0.20 dollars for cache reads and 2.50 dollars for cache writes. For comparison, Opus 5.5 is at 4 dollars input and 20 dollars output.

The saving therefore comes from consumption, not the rate. According to Anthropic, Sonnet 5.5 needs far fewer tokens to do the same work, which brings its cost per task up to 30% below Sonnet 5 in its tests. Anthropic adds that on several tests, Sonnet 5.5 at low or medium effort beats Sonnet 5's best score for about a tenth of the cost per task.

For a company, the consequence is direct: an agent's cost is measured per task handled, not by list price. A model that is more expensive per token can be cheaper per task if it uses fewer, and vice versa. That is why we recommend tracking cost per workflow and per agent, as proposed by our CADRE framework. The cost items of an agent in operation are detailed in what an AI agent costs in production.

Speed and effort level: a setting to decide

Sonnet 5.5 generates its answers more than 30% faster than Sonnet 5, making it, according to Anthropic, the fastest Sonnet to date. For an agent answering employees or customers, the gain shows immediately in waiting time.

Speed also depends on a setting often left at its default: the effort level. At low levels, the model answers faster and uses fewer tokens, which suits routine work. At high levels, it reasons longer and checks its work more thoroughly. Anthropic states that the default is "medium" in Claude Code and its apps, and "high" on the Claude Platform, the one developers use to build tools.

This detail matters. A team that tests the model in the app and then plugs it into a tool through the API is not comparing the same thing: the default setting differs. The effort level must be chosen explicitly, task by task, and recorded in the tool's documentation. It is a trade-off between quality, delay and cost, which is as much a business decision as a technical one.

What changes technically for a migration

Moving from Sonnet 5 to Sonnet 5.5 is more than changing a model name. The announcement flags several points worth checking before switching a tool in production.

• Thinking mode. If your tools run Sonnet with thinking turned off, Anthropic states that you need to switch to the new setting provided for this before migrating, and points to its migration guide.
• Cybersecurity safeguards. Because its cybersecurity capabilities have improved sharply, Sonnet 5.5 is the first Sonnet deployed with safeguards similar to Opus 5.5's. Routine software development is unaffected, but security requests deemed high risk visibly fall back to Sonnet 5. A security analysis tool must therefore be tested carefully.
• Thinking tied to the account. Anthropic has added protections against extraction of the model's reasoning. Most developers will not notice a change, but a tool that moves conversations between accounts must be checked.
• Data retention. The zero retention option remains available. If your contract depends on it, confirm that it applies to the new model in your configuration.

On the hosting side, availability at the three major cloud providers lets you keep your current infrastructure. The data location question is the same as with previous versions. We covered it in hosting an AI project in France.

Migrating an agent from Sonnet 5 to Sonnet 5.5: our method

On our projects, a model change always follows the same sequence, whatever the new model's reputation.

1. Replay the test set. Every agent we deliver comes with a series of real cases and their expected answer. We run them on the old and new model, at the same effort level, and compare results one by one. An average gain can hide a regression on an important case.

2. Measure cost and delay per task. We record tokens used and response time on the same cases. That is where the vendor's saving claim is confirmed, or not.

3. Check the business safeguards. What commits the company, a price, an availability, a reference, stays computed server side and checked before display, never produced by the model. A new model changes nothing to that rule, and we verify it still holds. We explained it in why an AI agent must never quote a price.

4. Switch gradually. Part of the traffic first, with a rollback possible within minutes, then all of it if the indicators hold. The old model stays available during the comparison.

5. Document. The model, its version, the effort level and the switch date are recorded. At the next change, you will know where you start from.

6. Tell the users. A model that writes differently, shorter or more direct, can surprise the people who use the tool every day. Anthropic states that Sonnet 5.5 writes more clearly than the previous generation. A short message to the teams, before the switch, explains what changes and whom to tell when an answer seems worse. This field feedback complements the test set, and it often arrives faster than the indicators. It also builds trust: people accept a change more easily when they know it is coming and that their feedback is used.

Combining Opus and Sonnet in one system

The close release of both models makes one architecture more attractive than before: giving each what it does best, within the same tool. Anthropic even quotes a tester who lets Opus set a project's architecture and Sonnet implement it.

In a business agent, this can take a simple form. The more capable model handles the steps that need judgment: understanding an ambiguous request, deciding how to handle it, reviewing a sensitive result. The faster, cheaper model handles repetitive, well-defined steps: extracting fields, rephrasing, classifying, producing a document from a template.

This split has a design cost. You must break down the process, define what passes from one step to the next, and test each step separately. It only pays off on real volumes. For a tool used a few times a day, a single well-tuned model is often the best choice. And for individual use in the office suite, the question is different, as we saw in Claude Opus 5.5 and GPT-6 Sol in Copilot.

Our takeaway

Sonnet 5.5 is an update worth testing quickly on tools running on Sonnet 5 today: same list price, fewer tokens used, faster answers. For many well-scoped agents, it should reduce cost per task and waiting time. But that promise is checked on your cases, not on a ranking.

For open-ended work that requires sustained judgment, Opus 5.5 keeps the edge, and Anthropic says so. The right reflex is not to pick a side, but to decide task by task, with measurements, and to record that choice so you can revisit it at the next model. Haiku 5.5, expected in the coming weeks, will raise the same question for the highest-volume tasks, and the same method will apply: a test set, cost and delay per task, a gradual switch.

If you already run tools on a Claude model, or wonder which one to choose for a first project, our AI diagnostic starts from your real processes. It tells you which model fits which task, at what cost, and what to test before deciding.

Sources

• Anthropic, Introducing Claude Sonnet 5.5, September 28, 2026
• Anthropic, Introducing Claude Opus 5.5, September 22, 2026

Frequently asked questions

Should we migrate all our tools from Sonnet 5 to Sonnet 5.5?

Not all at once. Start with the most used tools, replay their test set on both models and compare quality, cost and delay per task. Then switch gradually, keeping a rollback possible.

Can Sonnet 5.5 replace Opus 5.5 to cut costs?

On well-scoped, repetitive tasks, often. On open-ended work that requires sustained judgment, Anthropic itself says Opus 5.5 remains clearly stronger. The right answer depends on the task, and is checked on your cases.

Is the announced cost reduction guaranteed?

No. The per-token price is the same as Sonnet 5; the saving comes from lower consumption, measured by the vendor on its own tests. Only a measurement on your real tasks, at the effort level you will use, tells you what you will save.

Is our data handled differently with this new model?

The zero retention option is still offered, as for Sonnet 5 and Opus 5.5. Terms depend on your contract and the platform used, Claude Platform or cloud provider. Check it in your configuration before switching a sensitive tool.

When should we use Haiku rather than Sonnet?

Haiku 5.5 is not out yet; Anthropic announces it in the coming weeks, for very high-volume and cost-sensitive uses. Until then, Sonnet 5.5 at a low effort level is the cheapest option in the 5.5 family.