Claude Sonnet 5.5 is Anthropic’s mid-tier model, released on September 28, 2026. Token pricing stayed the same as Sonnet 5: $2 per 1 million input tokens and $10 per 1 million output tokens; the context window is 1 million tokens, and the maximum response is 128,000 tokens (Anthropic documentation). According to Anthropic, the model responds more than 30% faster than Sonnet 5 and costs up to 30% less per task thanks to fewer tokens and tool calls (Anthropic announcement). The main practical takeaway for business: in many agentic tasks, Sonnet 5.5 has come close to the flagship Opus 5.5 at half the token price, but if you are switching from Sonnet 5, you’ll need to update your code — the model has five breaking API changes.
This article is for executives, CIOs, and technical leads choosing a model for AI agents, chatbots, document processing, and software development. We cover capabilities, pricing, comparisons with neighboring models, API changes, availability for Russian companies, and scenarios where Sonnet 5.5 fits and where it doesn’t. A deep dive into the flagship is covered in the Opus 5.5 article; this text is not a substitute for a legal assessment of a specific project.
AIrassvet editorial team. Prepared on October 4, 2026 based on Anthropic’s documentation and announcement, verified on that date. The editorial team did not test the model independently: the test results are presented based on the vendor’s data and are labeled as such. The article was published on the AIrassvet website, which implements AI solutions for business; the services section appears before the conclusion. There are no affiliate links or compensation from Anthropic in this article.
Key takeaways
- Token pricing has not changed: $2 / $10 per 1 million tokens, cache read — $0.20, batch mode — 50% discount.
- The gains are in agent work: according to Anthropic’s tests, Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0 versus 10.3% for Sonnet 5 and 66.4% for Opus 5.5.
- Savings are per task, not per token: the model uses fewer tokens to do the same work; you need to measure the real benefit on your own data.
- Five breaking API changes: there is no
disabledmode for reasoning, no forced tool use, and new rules for reasoning blocks. - Russia is not included in the list of countries where Anthropic provides commercial access to the API and Claude.ai (Anthropic country list).
Contents
- What Claude Sonnet 5.5 is
- Specs and pricing
- What the model can do: benchmark results
- Comparison with Opus 5.5, Fable 5.1, and Haiku 4.5
- What a task costs: sample calculation
- What breaks when migrating from Sonnet 5
- When Sonnet 5.5 fits and when it doesn’t
- Availability in Russia and data requirements
- Safety and model refusals
- Frequently asked questions
- Primary sources and verification method
- How AIrassvet helps implement Claude Sonnet 5.5 or an alternative
- Bottom line: a strong midrange model, but don’t rush the migration
What Claude Sonnet 5.5 is
Claude Sonnet 5.5 is a language model in the Claude 5.5 lineup, which Anthropic positions as the “best balance of speed and intelligence.” In the lineup, it sits between the fast and affordable Haiku 4.5 and the more powerful Opus 5.5 and Fable 5.1. The model accepts text and images, responds in text, can call tools, control computers, and work with files and PDFs (model overview).
Anthropic describes the model’s main use cases as clearly defined everyday work: bug fixes and code review, creating documents, presentations, and spreadsheets, image design and recognition, and agentic tasks where execution reliability matters, such as research, verification, and drafting (Anthropic announcement).
Terms you’ll need later:
- Token — a unit of text used for model billing; in Russian text, one word usually takes several tokens.
- Context window — the amount of text the model considers in a single request: for Sonnet 5.5, it is 1 million tokens.
- Adaptive reasoning (adaptive thinking) — a mode in which the model decides for itself how long to “think” before answering; depth is set by the effort parameter.
- AI agent — software in which the model chooses the steps and calls tools itself (search, databases, APIs) until the task is complete.
Specs and pricing
Key parameters from Anthropic’s official documentation as of October 4, 2026:
| Parameter | Claude Sonnet 5.5 |
|---|---|
| Release date | September 28, 2026 |
| API identifier | claude-sonnet-5-5 (in Amazon Bedrock — anthropic.claude-sonnet-5-5) |
| Context window | 1 million tokens |
| Maximum response | 128,000 tokens (up to 300,000 in batch mode with a beta header) |
| Input / output | $2 / $10 per 1 million tokens |
| Cache writes for 5 minutes / 1 hour | $2.50 / $4 per 1 million tokens |
| Cache read | $0.20 per 1 million tokens |
| Batch Processing (Batch API) | 50% discount on input and output |
| Knowledge Cutoff | June 2026 |
| End of Support | no earlier than September 28, 2027 |
Source: the Claude Sonnet 5.5 model page.
Important detail — the tokenizer has not changed compared with Sonnet 5, so the same text produces the same number of tokens (what's new in Sonnet 5.5). That means the promised savings of up to 30% do not come from “compressing” the text, but from the model solving the task in fewer steps and responses. That is both a benefit and a caveat: for short prompts without tools, such as answering a single question, the cost difference may be almost zero.
What the model can do: benchmark results
Anthropic published a comparison of Sonnet 5.5 with its predecessor and with the flagship model. These are the vendor's numbers, not an independent audit:
| Benchmark (what it measures) | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic development in the terminal) | 70.6% | 10.3% | 66.4% |
| FrontierCode 1.1, main set (advanced coding) | 46.2% | 42.4% | 54.4% |
| CursorBench 4.0 (working in a code editor) | 55.5% | 34.1% | 57.8% |
| GDPval-AA v2.1 (office tasks for professionals, rating) | 1844 | 1449 | 1846 |
| OSWorld 2.1 (computer control) | 80.1% | 57.0% | 81.8% |
| Chartography (visual information recognition) | 61.6% | 15.6% | 64.4% |
Source: Claude Sonnet 5.5 announcement.
How to read the table. On agentic and office tasks, the gap between Sonnet 5.5 and Opus 5.5 has become small: in GDPval-AA the difference is 2 rating points, in OSWorld it's 1.7 percentage points, and in Terminal-Bench Sonnet 5.5 is even higher. On the hardest coding benchmark (FrontierCode), the flagship still leads by 8.2 percentage points. For business, that means the default choice of “we'll just use the strongest model” is now more expensive to justify: for common tasks, it makes sense to test Sonnet 5.5 first.
Another notable shift is computer control and image recognition. Growth from 57.0% to 80.1% in OSWorld and from 15.6% to 61.6% in Chartography makes the model a candidate for scenarios where an agent works with software interfaces, scans, and charts. But a benchmark is not your process: on real 1C forms or internal portals, the result needs separate validation.
Comparison with Opus 5.5, Fable 5.1, and Haiku 4.5
| Model | Context | Max. output | Input / output price per 1M | Speed (per Anthropic) | Knowledge Cutoff |
|---|---|---|---|---|---|
| Claude Fable 5.1 | 1M | 128K | $10 / $50 | slower | June 2026 |
| Claude Opus 5.5 | 1M | 128K | $4 / $20 | medium | June 2026 |
| Claude Sonnet 5.5 | 1M | 128K | $2 / $10 | fast | June 2026 |
| Claude Haiku 4.5 | 200K | 64K | $1 / $5 | fastest | February 2025 |
Source: Anthropic comparison table. Read more about the larger models in the reviews of Claude Opus 5.5 and Claude Fable 5.1.
For reference outside the Anthropic lineup: OpenAI’s GPT-6.1 Sol, introduced at DevDay 2026, costs the same $2 / $10 per 1M tokens for short-context requests (OpenAI pricing). GPT-6 Sol and Luna.
How much a task costs: example calculation
Below is an illustrative calculation based on public pricing, not a benchmark. Suppose an AI support assistant handles 1,000 requests, with each one using 3,000 input tokens (instruction, history, knowledge base excerpts) and 500 output tokens (response).
| Option | Input | Output | Total for 1,000 requests |
|---|---|---|---|
| Sonnet 5.5, standard mode | 3M × $2 = $6 | 0.5M × $10 = $5 | $11 |
| Sonnet 5.5, Batch API (-50%) | $3 | $2.5 | $5.5 |
| Opus 5.5, standard mode | 3M × $4 = $12 | 0.5M × $20 = $10 | $22 |
Calculation was done using the pricing from Anthropic’s model page as of October 4, 2026, excluding caching and taxes.
What changes the picture in practice:
- Caching. If the instruction and knowledge base repeat in every request, they can be cached: reading from cache costs $0.20 instead of $2 per 1M tokens. In Sonnet 5.5, the minimum cacheable portion was reduced to 512 tokens from 1,024 in Sonnet 5 (what's new).
- Agentic loops. In an agent, one task can mean dozens of model calls. Here, the claimed “up to 30% fewer tokens per task” has the biggest impact, but the cost also depends much more on how many steps the agent takes.
- Effort level. The depth of reasoning directly affects token count. By default, the API uses
high; Anthropic recommends starting withmediumfor well-defined agent tasks.
We covered how to keep costs under control as workloads grow in the article on keeping AI costs stable as token volume grows.
What breaks when moving from Sonnet 5
The most underappreciated part of the release is five incompatible API changes. Simply swapping the model ID to claude-sonnet-5-5 can cause 400 errors in a live product. According to Anthropic documentation (what's new in Sonnet 5.5):
| Change | What it was | What to do |
|---|---|---|
| Disabling reasoning | thinking: {"type": "disabled"} |
Pass {"type": "between_tools"} — works only when effort is no higher than high |
| Forced tool call | tool_choice with type any or tool |
Leave auto, enable strict tools mode (strict: true) or structured outputs |
| Reasoning blocks | It was possible to edit the conversation history | Keep history append-only; blocks are tied to the model, conversation, and account |
| Computer control | Tool computer_20251124 |
On the Claude API and Google Cloud, switch to computer_toolset_20260801 |
| Advisor tool | The advisor could be Sonnet 5 or Opus 4.8 | The advisor can only be a generation 5.5 model, Opus 5, Fable, or Mythos |
There are behavior changes without errors too. Text between tool calls now comes in reasoning blocks, and with the default settings its contents are empty — an interface that used to show users the agent’s intermediate comments will go silent without an error message. In addition, nonstandard values temperature, top_p and top_k return a 400 error: if your code has a hardcoded temperature of 0.2 somewhere, the request will stop working (model page).
Practical migration sequence:
- List every place where the code calls the model: parameters
thinking,tool_choice,temperature, computer control tools. - Fix the parameters according to the table above and update the SDK.
- Run 30–50 real tasks from the log on the old and new model and compare quality, token count, and time.
- Retune effort from scratch: Anthropic explicitly warns that the levels are calibrated differently and you cannot carry over the old setting.
- Shift traffic gradually, keeping the option to roll back to Sonnet 5.
When Sonnet 5.5 fits, and when it does not
Fits:
- AI agents with tools: intake triage, document preparation, data reconciliation across systems, knowledge base research;
- software development and code review, bug fixing, test generation;
- working with long documents — contracts, policies, technical documentation — thanks to the 1M-token window;
- high-volume batch processing where you do not need an immediate response: Batch API gives a 50% discount;
- controlling software interfaces, if the process has been pre-validated on real screens.
You should choose another model:
- Opus 5.5 or Fable 5.1 — for the most complex development and research tasks, where the cost of an error is higher than the token cost (according to FrontierCode, the gap remains significant);
- Haiku 4.5 or a specialized model — for simple classification and high-volume ticket routing; a generative model for choosing one option from a list is usually overkill;
- an on-premises model inside the company environment — if data cannot be sent to an external provider or you need guaranteed access without dependence on an overseas service. Cost comparison — in the article cloud or local LLM.
If a company has several types of tasks, it is more practical not to choose one model for everything, but to set up routing: simple requests to a cheaper model, complex ones to a stronger model. How to do this is described in the piece on AI model routing.
Availability in Russia and data requirements
Russia is not among the countries where Anthropic offers commercial access to the API and Claude.ai (Anthropic supported countries list). The model is also available through Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS (what’s new), but the terms of these platforms for Russian companies also need to be checked separately.
For companies with a foreign legal entity or subsidiary, access may be possible under the provider’s rules. For everyone else, it is more sensible to consider an architecture where the task is not tied to a single vendor: a unified model access layer and a fallback option — unified LLM API and AI service exit plan.
A separate issue is personal data. Sending it to an overseas service counts as cross-border data transfer and is regulated by Law No. 152-FZ; before sending customer data to any external model, you need to determine the legal basis, anonymize the data, or choose a model within your own environment. More details in the articles on 152-FZ and foreign AI and on transferring customer data to AI. This section is not legal advice.
Plus, for companies that do have access: Anthropic notes that zero data retention is available for Sonnet 5.5 (Anthropic announcement). Its terms are agreed with the provider separately.
Model safety and refusals
Sonnet 5.5 is the first model in the Sonnet lineup to receive cybersecurity safeguards at the Opus 5.5 level (Anthropic announcement). In the API, this looks like this: a rejected request returns HTTP 200 with stop_reason: "refusal" and the stop_details object, which specifies one of five categories — cyber, biology, competing model development, reasoning extraction, or other policy violations (documentation).
What this means for business system developers:
- the code must handle a refusal as a separate outcome, not as an empty response;
- for legitimate information security tasks (such as vulnerability analysis in your own code), it is worth checking in advance whether typical prompts fall under the restrictions;
- refusals count against request limits, and whether they are billed depends on the category.
Frequently asked questions
How much does Claude Sonnet 5.5 cost?
$2 per 1M input tokens and $10 per 1M output tokens — the same as Sonnet 5. Cache reads cost $0.20 per 1M tokens, and batch mode gives a 50% discount. The up to 30% savings Anthropic mentions refers to task cost, not token price.
How is Sonnet 5.5 different from Sonnet 5?
According to Anthropic, the model runs more than 30% faster, uses fewer tokens per task, and is significantly stronger in agentic coding, computer use, and image understanding. It also has five breaking API changes, so migration requires code updates.
Which should I choose: Sonnet 5.5 or Opus 5.5?
For typical agent and office tasks, it is worth testing Sonnet 5.5 first: in the vendor’s benchmarks, the gap to Opus 5.5 is small, while tokens are twice as cheap. Opus 5.5 makes sense for the most complex development and analysis tasks, where the cost of an error is higher than the token cost.
What is the context window for Claude Sonnet 5.5?
1M tokens in the input and up to 128K tokens in the response. In batch mode with a beta header, the response can reach 300K tokens.
Can Claude Sonnet 5.5 be used in Russia?
Russia is not on the list of countries where Anthropic provides commercial access to the API and Claude.ai. Companies without access should consider models available from Russia or deploying the model in their own environment.
Why do requests return error 400 after switching to Sonnet 5.5?
Most often because of the thinking: disabled parameter, forced tool invocation via tool_choice, nonstandard temperature, or the old computer use tool. All of these cases are covered in Anthropic’s migration guide.
Primary sources and verification method
The facts were verified on October 4, 2026, using the following sources:
- Claude Sonnet 5.5 — model page (Anthropic): specs, pricing, model comparison;
- What's new in Claude Sonnet 5.5 (Anthropic): API and behavior changes;
- Introducing Claude Sonnet 5.5 (Anthropic): benchmark results and positioning;
- Supported countries (Anthropic): country availability;
- OpenAI pricing: price of GPT-6.1 Sol for comparison.
Method: model information was taken from the documentation and the vendor announcement; benchmark results are those claimed by Anthropic and were not independently reproduced. The cost example was calculated using public pricing. Conclusions about model choice are editorial analysis, not the result of our own testing.
How AIrassvet helps implement Claude Sonnet 5.5 or an alternative
Choosing a model is only part of the job: you need to figure out which processes will benefit from an agent, how to connect it to your company systems, and what to do if the external model is unavailable or does not meet your data requirements. AIrassvet develops AI agents integrated with communication channels and business systems, and deploys language models, including on the client’s own servers. For the tasks in this article, three areas of work are especially useful:
- process and data audit: which steps can be delegated to the model, which data cannot be sent to an external service, and what class of model is needed;
- development of an AI agent MVP with integration into CRM, ERP, or 1C, quality testing on real tasks, and routing between models;
- deployment and fine-tuning of a language model within the company’s environment if cloud models are unavailable or do not meet security requirements.
The first step is to describe one process: its current metrics, data sources, constraints, and the criterion you will use to determine whether the model is performing well. After that, you can compare Sonnet 5.5 and alternatives on your own tasks, not someone else’s benchmarks. Discuss your project.
Bottom line: a strong mid-tier model, but don’t rush the switch
Claude Sonnet 5.5 is a model that, on agent and office tasks, has nearly caught up with the flagship Opus 5.5 in the vendor’s benchmarks, while keeping the $2 / $10 per 1M token price and a 1M token context window. For companies that have API access, this is a reason to rethink the default model choice: many tasks currently assigned to Opus can be moved to Sonnet 5.5.
But migration is an engineering task, not a one-line change: five breaking API changes, the new effort calibration, and a different format for intermediate messages all require testing on real data. Russian companies also need to consider the lack of official access and 152-FZ requirements. The next step is to choose one process, gather 30–50 typical tasks, and compare quality, time, and cost across several models, including an on-premises option.