Claude Sonnet 5.5 explained: same token price, cheaper cache reads, one effort trap

Anthropic released Claude Sonnet 5.5 on 28 September 2026 at the same price as the model it replaces: $2 per million input tokens and $10 per million output. Anthropic says it writes more than 30 per cent faster than Claude Sonnet 5, needs far fewer tokens for the same job, and costs up to 30 per cent less per task as a result.

For most businesses this is the Claude model you will use every day, and the trade is a good one. There is one catch, and it sits in the effort setting. Above high effort, Claude Opus 5.5 gets the same or better scores for less: at xhigh it matches Sonnet 5.5’s best score on the Artificial Analysis index for $3.46 per task against $5.46 (checked 19 October 2026). Leave it at medium or high and you get the upgrade at the old price.

The answer in brief

• Claude Sonnet 5.5 is Anthropic’s mid-tier model, released on 28 September 2026 as the second model in the Claude 5.5 family, after Opus 5.5. Claude Haiku 5.5 followed on 7 October 2026.
• It costs $2 per million input tokens and $10 per million output, unchanged from Sonnet 5. Cache reads are $0.10 (Anthropic halved them on 7 October 2026) and five-minute cache writes $2.50.
• Anthropic says it generates output more than 30 per cent faster than Sonnet 5 and costs up to 30 per cent less per task, because it needs far fewer tokens for the same work. Early testers also reported fewer steps and tool calls.
• On the Artificial Analysis Intelligence Index it scores 56 at maximum effort, second only to Opus 5.5 on 58, and 18 points above Sonnet 5 (checked 19 October 2026).
• The effort setting decides the bill. At high effort it beats Sonnet 5’s best score for about a sixth of the cost per task. Above high, Opus 5.5 is the better buy: at xhigh it matches Sonnet 5.5’s maximum score of 56 for $3.46 per task against $5.46.
• Migrating an API integration from Sonnet 5 needs code changes: disabled thinking, forced tool use and the old computer-use tool all return errors, and edited conversation history fails on newer API accounts.
• Try it on one everyday workflow at medium effort, compare tokens, steps and rework against Sonnet 5, then move traffic in stages. Keep Opus 5.5 for long, open-ended work.

What is Claude Sonnet 5.5, and where can you use it?

Anthropic positions Sonnet 5.5 as the faster, lower-cost complement to Opus 5.5. Opus 5.5 is built for complex work that needs sustained judgement. Sonnet 5.5 is for well-scoped everyday tasks: fixing bugs, drafting documents, building slides and spreadsheets, and iterating quickly on smaller jobs. Anthropic also says early testers noticed a sharper eye for design and clearer writing than the previous generation.

It takes text and images and returns text. Adaptive thinking is on by default and you steer its depth with the effort setting, which has five levels: low, medium, high, xhigh and max. The reliable knowledge cutoff is June 2026. Anthropic’s deprecation schedule says it will not retire the model sooner than 28 September 2027, and Sonnet 5 stays active until at least 30 June 2027, so nobody is being forced to move.

Claude Sonnet 5.5 at a glance

Where you can use it

Anthropic lists Sonnet 5.5 on the Claude Platform, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, with zero data retention available as on Opus 5.5 and Sonnet 5. Its launch page says the model is in Claude Code and the Claude apps, where the default effort is medium. Anthropic’s pages do not say which plans get it as the default model, so check your model picker rather than assuming.

As with Opus 5.5, a Claude subscription does not mean your own application uses the API model. Check the product you intend to use, and review zero-data-retention and processor terms under UK GDPR before client data goes through it. Anthropic bills in US dollars, so convert at your own rate and add VAT where it applies.

How much does Claude Sonnet 5.5 cost?

Per million tokens, from Anthropic’s pricing page (checked 9 October 2026). Cache write rates depend on the cache lifetime, so both are shown:

Input and output prices have not moved. Sonnet 5 launched at $2 and $10 as an introductory rate, Anthropic made that permanent on 10 August 2026, and Sonnet 5.5 inherits it. Cache reads are now cheaper: Anthropic halved Sonnet 5.5’s to $0.10 per million tokens on 7 October 2026. GPT-6 Sol sits at the same $2 and $10 on OpenAI’s API, so the price comparison between the two mid-tier models comes down to tokens per task, not the rate card.

That is where Anthropic’s saving claim lives. Its own testing puts the cost per task at up to 30 per cent below Sonnet 5, because the model finishes the same work in fewer tokens, and early testers saw fewer tool calls too. The customer figures on the launch page point the same way: Slack reports about 14 per cent fewer output tokens on its Slackbot evaluations, Box 12 per cent fewer total tokens than its previous model, Lovable a third fewer tool calls, and Balyasny Asset Management about 121,000 tokens per answer on its finance tasks against 497,000 for Sonnet 5. Those are vendor-selected testimonials, and none of them is a marketing workload, so measure your own.

Two smaller notes for budgets. Cache writes at the one-hour rate are $4 per million tokens, not $2.50, so agents that keep a long-lived cache pay more up front. And on the Batch API, Sonnet 5.5 costs $1 input and $5 output, which is the Haiku 4.5 list price for a much stronger model, if the job can wait.

Which effort setting should you use with Claude Sonnet 5.5?

Medium for routine work and high for anything that needs checking. Avoid xhigh and max unless a side-by-side test proves they earn their cost. The independent numbers make the case.

Artificial Analysis ran Sonnet 5.5 at every effort level on its Intelligence Index v4.3.2, a composite of ten evaluations, with Anthropic’s default fallback switched on. Its cost per task for Sonnet 5.5 has fallen at every setting since launch week; at max it was $7.62 on 2 October. Cost per task is Artificial Analysis’s weighted average across the index, at list prices:

Read it from the bottom up. Sonnet 5.5 at medium beats Sonnet 5’s best score for about a tenth of the cost per task. At high it scores nine points above Sonnet 5’s best for about a sixth of the cost. That is the upgrade, and it arrives without a price change.

Then look at the top. Above high, Opus 5.5 is the better buy. Sonnet 5.5 at xhigh scores 52 for $2.01 per task, while Opus 5.5 at high scores 54 for $1.82. Sonnet 5.5 at max reaches 56 for $5.46, a score Opus 5.5 matches at xhigh for $3.46. Opus 5.5 at max costs more, at $5.98, but scores 58. When a job needs more than Sonnet 5.5 at high, move it to Opus 5.5 rather than turning the effort up. Artificial Analysis no longer lists GPT-6 Sol; GPT-6.1 Sol scores 50 for $0.32 per task at high effort, ahead of Sonnet 5.5 at high on both score and cost.

Anthropic’s own charts say the same thing in its own words: Sonnet 5.5 complements Opus 5.5 best at lower effort settings, and at Low or Medium it beats Sonnet 5’s best score on several benchmarks for about a tenth of the cost per task. Its FrontierCode footnote even records Sonnet 5.5 scoring lower at Max than at Xhigh, because at Max it more often ran a multi-agent code-review skill that overran the task.

Two caveats. At launch, Artificial Analysis ran these evaluations on a pre-release deployment that carried the structured-outputs bug Anthropic has since fixed, and said it would re-run the relevant ones, so check its live pages before you rely on the decimals. And Anthropic’s own prompting advice is to start at high, the API default, unless the workload is agentic or latency-sensitive; our medium-first recommendation is the cheaper starting point for the routine marketing jobs this post is about. For harder work, take Anthropic’s advice and start at high.

One migration detail matters here too. Anthropic’s what’s new page says the effort levels are recalibrated, so a setting does not produce the same amount of thinking as it did on Sonnet 5. Re-run your effort sweep rather than carrying a value over.

For the prompt lines Anthropic has tested on Sonnet 5.5, and the old habits to delete, see our guide to how to prompt Claude.

How does Sonnet 5.5 compare with Sonnet 5 and Opus 5.5?

Anthropic’s launch figures show a large jump over Sonnet 5 and near-parity with Opus 5.5 on several evaluations. They are vendor numbers, and Anthropic itself adds that in its own and external testing Opus 5.5 remains clearly stronger at complex, open-ended work. Both statements are true at once.

Two caveats come from Anthropic’s own footnotes. The GDPval-AA and AA-Briefcase scores were run by Artificial Analysis on a pre-release deployment with a bug that could degrade structured outputs, since fixed, which Anthropic expects to have understated Sonnet 5.5 slightly if it had any effect. And the Opus 5.5 Terminal-Bench figure is its xhigh score, its best. The launch page does not say which setting the Sonnet 5.5 figure used, but the system card does (checked 9 October 2026): maximum effort, run in Claude Code, with Opus 5.5 scoring 64.8 per cent at that setting. Anthropic puts the standard error at about 2.5 points for each model, so a lead of four to six points is not much bigger than the noise.

Artificial Analysis’s independent runs are more mixed than the vendor table. On its Terminal-Bench 4.0 run Sonnet 5.5 scored 64 per cent against 60 per cent for Opus 5.5 and GPT-6 Astra, and it matched Opus 5.5 on AA-Briefcase, GDPval-AA and AutomationBench-AA, with the heavy token use noted above. On factual knowledge it lags: 54 per cent accuracy on AA-Omniscience against 66 per cent for Opus 5.5, though with a lower hallucination rate of 47 per cent against 59. It also sits about six points behind Opus 5.5 on Humanity’s Last Exam and SciCode. For a marketing team, that pattern suggests Sonnet 5.5 for drafting, formatting and tool-driven work, and Opus 5.5 where the facts have to be right first time.

Anthropic also says Sonnet 5.5 is the first Sonnet to finish Pokémon Red working only from screenshots, and that two experts judged its first draft of a 10-slide operating review, built from a public company’s earnings materials and a slide template, ready to send. Treat both as a test list for your own trial, not a forecast.

What breaks if you migrate from Sonnet 5?

Chat users change the model picker and carry on. API integrations need a proper check, because Anthropic’s migration guide lists what changes for code coming from Sonnet 5:

1. Thinking cannot be disabled. thinking: {"type": "disabled"} is rejected. The lowest setting is now between_tools, which turns off up-front thinking but still returns progress notes between tool calls as thinking blocks. It works at low, medium and high effort only; at xhigh or max you must use adaptive thinking.
2. Forced tool use returns an error. tool_choice values of any or a named tool are rejected. Use auto with strict tools and a clear instruction in the prompt.
3. The old computer-use tool is rejected on the Claude API and Google Cloud. Move computer_20251124 to computer_toolset_20260801. Amazon Bedrock keeps the old tool.
4. Thinking blocks are tied to the model, the conversation and the account. Sonnet 5.5 reads thinking carried over from Sonnet 5, Opus 4.8 and Haiku 4.5, but not from Opus 5, Opus 5.5 or Fable, and a request that routes back down to Haiku 4.5 drops Sonnet 5.5’s blocks. For API accounts created on or after 31 August 2026, editing the system prompt, tools or earlier messages after a Sonnet 5.5 thinking block makes the request fail by default. Keep conversations append-only.
5. Advisor pairings change. As the executor, Sonnet 5.5 accepts advice only from Mythos, Fable, Opus 5.5, Opus 5 or itself. Opus 4.8, Opus 4.7 and Sonnet 5 advisors return an error.
6. Longer notes between tool calls now arrive in thinking blocks. Anything beyond a sentence or two moves out of the text stream, and those blocks are empty at the default display setting, so agent progress updates go quiet unless you change it.

Non-default temperature, top_p or top_k values, thinking budgets and assistant prefill also return errors, but they already did on Sonnet 5, so code that has been running there is clear of them. Anthropic’s Claude Code skill can apply most of the changes above across a code base with /claude-api migrate this project to claude-sonnet-5-5. It still produces a checklist for you to verify by hand, and the effort recalibration is the item most likely to change your bill.

Safeguards and fallbacks

Sonnet 5.5’s cyber capabilities are, in Anthropic’s words, comparable to Opus 5’s, so it is the first Sonnet to launch with the cyber safeguards and fallbacks built for the larger models. Routine software work is unaffected, but higher-risk cybersecurity requests visibly fall back to Sonnet 5. Artificial Analysis saw that fallback fire in about 0.1 per cent of its index tasks, mostly in Terminal-Bench, and fall back to Sonnet 5 every time. Biology safeguards are the same as Sonnet 5’s. Anthropic says its expanded Cyber Verification Program will open to vetted defenders soon, and the Life Sciences Verification Program is already open to organisations.

Sonnet 5.5 is also the first Sonnet to ship with classifiers that block reasoning extraction, Anthropic’s response to distillation attacks that use thousands of fake accounts to copy a model’s capabilities. Most developers will not notice. If you move conversations between accounts, including switching accounts mid-session in Claude Code, read the preserved-thinking notes in the docs first. The system card has the full alignment and safeguards detail.

What should UK marketing teams test first?

Pick the everyday workflow that runs on Sonnet 5 today and give Sonnet 5.5 the same brief, the same files and medium effort. Record tokens in, tokens out, tool calls, minutes to finish and minutes of human rework. Then run the same job at high and compare. Do not start at max.

Good first tests for a Manchester or Leeds team:

• a monthly report deck built from exported data and your own slide template, because Anthropic is making a specific claim about slides
• a support or sales inbox triage that already runs on Sonnet 5, where Zendesk’s 20 per cent faster ticket handling against its production Claude models is the figure to beat
• a site audit that rereads the same brand rules on every turn, where fewer tool calls and cache reads at $0.10 decide the cost
• a fact-heavy brief that must cite primary pages, to see where the AA-Omniscience gap shows up in your work

Treat what comes back as a draft from a capable new starter. Review it, check the citations, and keep a named person accountable for anything that publishes.

Should your business switch to Claude Sonnet 5.5?

For work already on Sonnet 5, yes, after a measured trial, and sooner than you would for an Opus upgrade. The price is unchanged, the speed and token savings are the kind that show up on an invoice, and Sonnet 5 stays available until at least June 2027 if something breaks.

Keep effort at medium or high. Move a job to Opus 5.5 when it needs sustained judgement or has to be right on the facts, and keep Fable 5.1 for the long, expensive runs that Opus 5.5 fails. If GPT-6 Sol or GPT-6 Astra is already in the stack, run it on the same task at the same cost per task. No model wins every job.

If your team wants a measured trial plan for Sonnet 5.5, talk to AIWIZ.