Claude Sonnet 5.5

Claude Sonnet 5.5 — Text Generation (LLM)

What is Claude Sonnet 5.5?

Claude Sonnet 5.5 is Anthropic's most agentic Sonnet model, built for well-scoped everyday engineering and knowledge work. Announced in September 2026 as the second model in the Claude 5.5 family, it brings near-Opus coding, debugging, and reasoning to the Sonnet tier, paired with a 1M-token context window and fine-grained control over how hard the model thinks. It runs as a synchronous text-generation API: send a prompt, optionally add an image or a system instruction, and receive the answer directly.

Key Features

• •Agentic coding across the whole software lifecycle: scoping changes, building features, fixing bugs, and reviewing code.
• •Adaptive thinking through five effort levels (low, medium, high, xhigh, max), so you trade reasoning depth for speed on demand.
• •Near-Opus benchmark quality, far ahead of Sonnet 5 on agentic coding and knowledge-work evaluations.
• •A 1M-token context window for large codebases and long documents.
• •Optional image input for multimodal analysis and structured JSON-Schema output for machine-readable results.

Best Use Cases

Sonnet 5.5 is strongest at debugging and code review. In testing it reliably found multiple planted bugs in a Python function, named a root cause for each, returned a corrected version, and generated a pytest suite that failed on the original and passed on the fix. It is equally capable at drafting clear documents, summaries, and structured data, and at well-defined agent tasks such as investigation, triage, and drafting. For ambiguous, open-ended problems that need sustained judgment, reach for Opus 5.5 instead.

Prompt Tips and Output Quality

Be specific: state the task, the inputs, and the exact output you want, whether that is corrected code, tests, or a JSON shape. Raise effort to high or xhigh for coding and multi-step analysis, and keep it low for chat and quick extraction. Add a system instruction to lock format or persona, and set response_format when you need machine-readable JSON.

How to Use Claude Sonnet 5.5

Claude Sonnet 5.5 is a synchronous text API: send a prompt and read the answer directly. The prompt is the only required field; every other parameter tunes how the model thinks and what it returns.

Recommended Settings

• •Coding, debugging, and multi-step analysis: set effort to high or xhigh. In testing at high, Sonnet 5.5 engaged its reasoning before answering and produced careful, checkable results.
• •Chat, extraction, and classification: keep effort at low for the fastest responses.
• •Correctness-critical work: use max when a right answer matters more than latency.
• •System behaviour: pass instruction to fix tone, persona, or output rules for the whole request.
• •Multimodal: attach an image (PNG, JPEG, WebP, or GIF) only when the answer depends on visual input.
• •Structured output: supply a JSON Schema via response_format for machine-readable results — list every property in required and set additionalProperties to false.

Example prompt to try

This prompt ran end to end at effort: high during testing:

"This scheduling helper is meant to coalesce a day's busy time-blocks — each given as a [start_minute, end_minute] list — into the fewest non-overlapping blocks, but it returns wrong results on several inputs. [...] Give a one-line root cause for each bug, return a corrected coalesce_blocks, and write a pytest suite that fails on the original and passes on your fix — include a touching-endpoint case, an unsorted-input case, and a case that checks the input list is left unchanged. Finish with the Big-O time complexity of your corrected version."

The model named a root cause for each of the three planted bugs (a strict < that should be <=, missing input sorting, and in-place mutation of the caller's list), caught a bonus empty-input edge case, returned a corrected function, wrote a multi-case pytest suite, and added an O(n log n) complexity note. Tip: paste the full buggy function inside a fenced code block and name the output you want (corrected code plus tests) so the response drops straight into your project.

Further prompts it is built for, phrased as suggestions to adapt: "review this pull request and flag risky changes", "refactor this module without changing the public API", "extract these fields from the text as JSON".

When should I use Sonnet 5.5 over Opus 5.5?

Use Sonnet 5.5 for well-scoped everyday coding and knowledge work; use Opus 5.5 for ambiguous work that needs careful judgment.

Does it accept images?

Yes. Add a PNG, JPEG, WebP, or GIF for multimodal analysis; video input is not supported.

How do I control reasoning depth?

Set the effort parameter from low through max, higher for harder tasks.

Can it return structured output?

Yes. Supply a JSON Schema via response_format; json_object mode is not supported on Claude.

What context length does it handle?

Up to 1M tokens, enough for large codebases and long documents.