Claude Sonnet 5.5 was released on 28 September 2026, and its API ID is claude-sonnet-5-5. Anthropic sells it at $2 per million input tokens and $10 per million output tokens. On Claudexia on 19 November 2026 the same model costs $0.62 and $2.92 on the Premium channel and $0.54 and $2.00 on the Cheap channel. Claude Code needs version 2.1.284 or newer, and on paid plans and the Anthropic API its default is Opus 5.5, not Sonnet.
What this model is
Sonnet 5.5 is the second model of the Claude 5.5 family: Opus 5.5 came out on 22 September and Haiku 5.5 on 7 October (Anthropic release notes). On the model page Anthropic calls it the best combination of speed and intelligence and lists its jobs: everyday coding, agents, enterprise workloads, data analysis, writing, and image understanding. Its status is "Active (latest)", and retirement is not before 28 September 2027.
In the 28 September announcement Anthropic splits the roles this way. Sonnet 5.5 is strongest on well-scoped everyday tasks, bug fixes, and polished documents, spreadsheets and slides. Opus 5.5 stays clearly stronger on complex, open-ended work that needs sustained judgment. That is the vendor's assessment, and we have not compared the models ourselves.
API parameters
| Parameter | Value |
|---|---|
| ID in the Claude API | claude-sonnet-5-5 |
| Context | 1M tokens |
| Maximum output | 128K tokens |
| Input and output | text and images in, text out |
| Thinking | adaptive, default effort high |
| Knowledge cutoff | June 2026 |
| Minimum cacheable prompt | 512 tokens |
| Retirement | not before 28 September 2027 |
Setting temperature, top_p or top_k returns a 400 error on this model, as our article on the 400 error from temperature, top_p, top_k explains. Adaptive thinking is on by default, and thinking: {"type": "disabled"} returns 400. The replacement is the value between_tools, covered below.
The Claudexia catalog lists an output limit of 32,000 tokens for claude-sonnet-5-5, while Anthropic documents 128K. For claude-opus-5-5 the catalog says 128,000. We did not check how much output the gateway actually lets through for Sonnet, so treat 32,000 as the ceiling for long answers.
Price at Anthropic and on both Claudexia channels
Prices per 1M tokens. Anthropic: the pricing page. Claudexia: the live price list on 19 November 2026.
| Per 1M tokens | Anthropic | Claudexia Premium | Claudexia Cheap |
|---|---|---|---|
| Input | $2 | $0.62 | $0.54 |
| Output | $10 | $2.92 | $2.00 |
| Cache read | $0.10 | $0.08 | $0.10 |
| Cache write (5 minutes at Anthropic) | $2.50 | $0.72 | $0.24 |
Claudexia is below the list price. Premium input is 31% of $2 and Cheap input is 27%. For output, Premium is 29% of $10 and Cheap is 20%. That is one number divided by another from the table, and the price list does not say why it is set this way. Anthropic also offers a 50% discount in the Batch API and a one-hour cache write at $4 per million. Claudexia has no Batch API. Prices change, so open the catalog on the models page before you calculate.
Which channel to pick
Cheap is cheaper on input by $0.08 per million, on output by $0.92, and on cache writes by $0.48. Premium is cheaper only on cache reads, by $0.02 per million. So Premium wins only where almost all of the input is read from cache. Here are two hypothetical agent turns (arithmetic, not a measurement of real sessions):
| Agent turn | Premium | Cheap | Anthropic |
|---|---|---|---|
| 150,000 tokens from cache, 5,000 cache write, 500 fresh input, 1,500 output | $0.0203 | $0.0195 | $0.0435 |
| 300,000 tokens from cache, 500 each for cache write, fresh input and output | $0.0261 | $0.0314 | $0.0373 |
Each API key has its own channel: Premium by default, Cheap, or both with switching (auto_premium, auto_cheap) if switching is enabled for the account. You pay the rate of the channel that answered, and the rules are in the channels documentation. Cheap can be closed for an account, and then a request returns 403 channel_not_allowed. For ways to cut cost, see the article on prompt caching, and for a price roundup of all Claude models see Claude API pricing in 2026.
Claude Code: version, default model, setup
Client version
Sonnet 5.5 requires Claude Code 2.1.284 or newer, Opus 5.5 requires 2.1.280, and Haiku 5.5 requires 2.1.293 (model configuration docs for Claude Code). Check claude --version, and update with claude update. If a request for the new model fails from an older version, the documentation points to its section "Claude Code does not support this model".
What is the default
According to the documentation on 19 November 2026, the default on Pro, Max, Ultra, Team, Enterprise and the Anthropic API is Opus 5.5. Before version 2.1.280 the default on Pro and Team Standard was Sonnet 5. Sonnet 5.5 is selected by the sonnet alias on the Anthropic API, and opusplan mode switches to it after planning: Opus builds the plan, Sonnet executes. The default effort in Claude Code is medium, while the API default is high. In Claude Code, thinking cannot be turned off on Sonnet 5.5.
Setup through Claudexia
The address has no /v1, and the key starts with sk_cdx_. All three IDs below exist in both Premium and Cheap, so the configuration fits a key on either channel.
{
"env": {
"ANTHROPIC_BASE_URL": "https://api.claudexia.tech",
"ANTHROPIC_AUTH_TOKEN": "sk_cdx_YOUR_KEY",
"ANTHROPIC_MODEL": "claude-sonnet-5-5",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5-5",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5-5",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-5-5"
}
}
The file ~/.claude/settings.json sits in your home folder. Do not put the key into .claude/settings.json inside a project, because it would end up in the repository. You do not need the [1m] suffix: Sonnet 5.5 has a native 1M window, and according to the documentation Claude Code recognizes such a model by an ID that contains its API ID claude-sonnet-5-5. We copied this configuration from our own guides and did not run it through Claudexia while preparing this article. Step-by-step installation is in the Claude Code setup guide, and the November versions are covered in Claude Code in November 2026.
Cybersecurity refusals
Sonnet 5.5 is the first Sonnet model released with cybersecurity safeguards. A request that trips a safeguard comes back with HTTP 200 and stop_reason: "refusal". In Claude Code v2.1.219 or later on the Anthropic API, a request flagged as cybersecurity re-runs on Sonnet 5 by itself, and you set no variable for that. The session then continues on Sonnet 5, and /model takes you back to Sonnet 5.5. If you pick "Ask each time" in /config (switchModelsOnFlag: false in the settings file), Claude Code asks before the switch, and in an interactive session it may also ask on the first such request whether to switch automatically from then on. If availableModels blocks Sonnet 5, no fallback happens and the request ends with a refusal. A request flagged as biology has no fallback model and ends with a refusal. The Claude Code documentation sets the ANTHROPIC_DEFAULT_SONNET_MODEL condition only for Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry: those providers use their own model IDs, and a pin that names Sonnet 5.5 itself leaves the refusal standing. How Claude Code behaves with the Claudexia address and the claude-sonnet-5-5 entry from the example above, we did not check. In your own code, handle the refusal through stop_reason: Anthropic has server-side fallback (in beta), and we did not try it through Claudexia.
How Sonnet 5.5 differs from Sonnet 5
| Parameter | Sonnet 5 | Sonnet 5.5 |
|---|---|---|
| Release date | 30 June 2026 | 28 September 2026 |
| Status at Anthropic | Legacy | Active (latest) |
| Knowledge cutoff | January 2026 | June 2026 |
| Price per 1M, input / output | $2 / $10 | $2 / $10 |
| Cache read per 1M | $0.20 | $0.10 |
| Minimum cacheable prompt | 1,024 tokens | 512 tokens |
Both use the same tokenizer, so identical text gives identical token counts. Anthropic says Sonnet 5.5 generates output 30% or more faster than Sonnet 5 and, in its tests, costs up to 30% less per task at the same price per token. From the announcement table: on Terminal-Bench 4.0 it scores 70.6% against 10.3% for Sonnet 5, and on CursorBench 4.0 55.5% against 34.1% (Opus 5.5 scores 57.8%). All of these are Anthropic's numbers, and we did not reproduce them.
For code written against Sonnet 5, Anthropic's documentation lists changes that can break it:
- Switching off thinking before the answer now works only with
thinking: {"type": "between_tools"}at efforthighor lower. The value"disabled"returns 400, andbetween_toolsalso returns 400 atxhighandmax. - Forced tool use (
tool_choicetypesanyandtool) returns 400.autoandnonework. - Thinking blocks are tied to the model and the account, and a conversation has to stay append-only: editing the system prompt, tools or early messages before you resend a block can produce a 400.
- The old
computer_20251124tool is not accepted on the Claude API, andcomputer_toolset_20260801is required. - The advisor tool accepts only some advisors: several models, Sonnet 5 among them, are rejected in that role.
- Notes between tool calls that run longer than a sentence or two arrive in thinking blocks, while shorter remarks stay plain text. An app that streams those notes to the user goes silent between calls until you set
displayor choosebetween_tools.
That is Anthropic's API behavior according to its documentation. How Claudexia handles such requests, we did not check. In Premium claude-sonnet-5 and claude-sonnet-5-5 cost the same, $0.62 / $2.92, so moving does not change the token price, but claude-sonnet-5 exists only there and not in Cheap. If you are used to Sonnet 5, read our overview Claude Sonnet 5: the workhorse versus Opus as well.
Sonnet 5.5 or Opus 5.5
At Anthropic, Opus 5.5 costs $4 and $20, exactly twice Sonnet 5.5. The gap on Claudexia is different. In Premium, Opus 5.5 costs $0.88 and $4.40, which is 1.4 times more on input and 1.5 times more on output. In Cheap it is $1.06 and $4.00, nearly double on both numbers (1.96 and 2.0). The catalog output limit is 128,000 for Opus and 32,000 for Sonnet.
Claude Code has an opusplan mode: Opus plans and Sonnet executes. For other cases we would use Sonnet 5.5 on well-described tasks and move to Opus when the task is open-ended or the agent repeats mistakes even with clean context. The choice is covered in more detail in the comparison of Opus 5.5, Sonnet 5.5 and Haiku 5.5, and Opus has its own Russian article, "Claude Opus 5.5: free or buy". Compare on your own tasks: the 20-minute method is in Public benchmarks miss your task.
Example call
Anthropic SDK, address without /v1. The example has no temperature, top_p or top_k, since they return 400 on this model.
import os
from anthropic import Anthropic
client = Anthropic(
api_key=os.environ["CLAUDEXIA_API_KEY"],
base_url="https://api.claudexia.tech",
)
message = client.messages.create(
model="claude-sonnet-5-5",
max_tokens=1024,
messages=[{"role": "user", "content": "Split this PR into logical commits."}],
)
text = "".join(block.text for block in message.content if block.type == "text")
print(text or f"No text, stop_reason: {message.stop_reason}")
On this model, a thinking block with an empty thinking field can come before the text: display defaults to omitted. That block has no text field, so message.content[0].text fails, and you collect the text from blocks of type text.
Signing up at claudexia.tech takes a login and password, or Google, GitHub or Telegram, and there is no email field. The minimum top-up is $1, paid in rubles through SBP (Russia's Faster Payments System), with a Russian card, or in crypto. A key on the Premium channel opens both Claude and GPT models, for example GPT-6 Sol and Luna.
FAQ
When was Claude Sonnet 5.5 released?
On 28 September 2026, as both the model page and the Anthropic release notes confirm. Haiku 5.5 followed on 7 October.
What is the ID of Sonnet 5.5?
claude-sonnet-5-5 in the Claude API and on Claudexia. On Amazon Bedrock the ID differs: anthropic.claude-sonnet-5-5.
Is Sonnet 5.5 the default in Claude Code?
No. According to the documentation on 19 November 2026, the default on paid plans and on the API is Opus 5.5. You select Sonnet 5.5 with the sonnet alias, the /model command, or the ANTHROPIC_MODEL variable.
Do I need the [1m] suffix?
No. Sonnet 5.5 has a native 1M window, and Claude Code recognizes it by the ID claude-sonnet-5-5.
Is Sonnet 5.5 on the Cheap channel?
Yes. Cheap has three Claude models: Haiku 5.5, Opus 5.5 and Sonnet 5.5. Sonnet 5 exists only in Premium.