Claude Sonnet 4.5 retirement date: November 30, 2026

Claude Sonnet 4.5 vs Claude Sonnet 5.5: price and limits

Claude Sonnet 5.5 uses Anthropic's newer tokenizer, which produces about 30% more tokens for the same text (Anthropic token counting docs). Recount with the Claude token counter before comparing costs.

Prices and limits from models.dev (checked 2026-11-02). Confirm on the official pricing page. Compare the cost for your traffic in the old vs new model calculator.

Migrating from Claude Sonnet 4.5 to Claude Sonnet 5.5

Draft: not yet reviewed. These notes are drafted from the official migration guide and haven't been reviewed by a human yet. Check each item against its source.

Official guide: https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide

Breaking changes (requests that will error)

• Assistant prefill returns a 400 error. Claude Sonnet 5.5 rejects a prefilled last assistant turn. Use structured outputs (or tools with enum fields) for output format, the system prompt for preambles, and the user turn for continuations and context reminders. source
• Thinking budgets (budget_tokens) return a 400 error. Sonnet 4.5 uses thinking: {type: "enabled", budget_tokens: N}. Remove the budget and set output_config.effort instead; there’s no fixed mapping, so evaluate at two or three effort levels. source
• temperature, top_p and top_k return a 400 error. Remove sampling parameters and steer behavior with prompting. source
• Forced tool use (tool_choice of any or tool) returns a 400 error. Send tool_choice: {type: "auto"}, mark the tool strict: true, and say in the prompt when to use it. Also applies to the token counting endpoint. source
• thinking: {type: "disabled"} returns a 400 error. To run without up-front thinking, send thinking: {type: "between_tools"} (accepted at low, medium and high effort only). source
• Computer use: computer_20250124 is not accepted. On the Claude API use the computer_toolset_20260801 toolset; on Amazon Bedrock use computer_20251124. Remove the fine-grained-tool-streaming-2025-05-14 header when you move to the toolset. source
• output_format without its beta header returns a 400 error. Move structured output settings to output_config.format. source

Behavior changes (no error, different output)

• Thinking runs by default. A request with no thinking field now runs with adaptive thinking. Read content blocks by type (not content[0].text), pass thinking blocks back unchanged in tool loops, and revisit max_tokens, which covers thinking plus text. source
• About 30% more tokens for the same text. Sonnet 5.5 uses the newer tokenizer. Recount prompts with the token counting endpoint and re-check max_tokens, context fit and cost. source
• Images can cost about 2.5x more tokens. Sonnet 5.5 uses the high-resolution image tier (up to 2,576 px on the long edge and 4,784 visual tokens per image). source
• Text between tool calls can arrive as thinking blocks. Longer progress notes between tool calls come back as thinking blocks, empty at the default display, so UIs that show them go quiet unless you set display. source
• More refusal categories. Declines return stop_reason: "refusal" with a category in stop_details. Handle refusals and consider server-side fallback. source
• Tool input escaping can differ. Parse tool call input with a standard JSON parser rather than string handling. source

Checklist

• Search your code, config and env vars for claude-sonnet-4-5-20250929 and the claude-sonnet-4-5 alias, and switch to claude-sonnet-5-5
• Re-run your eval set against claude-sonnet-5-5 before switching traffic
• Replace assistant prefills, thinking budgets, sampling parameters and forced tool_choice
• Set output_config.effort explicitly (Sonnet 4.5 had no effort parameter) and re-run an effort sweep
• Recount tokens: Claude 4.7+ models use a newer tokenizer that produces about 30% more tokens for the same text
• Remove interleaved-thinking-2025-05-14 and any context-window beta header; replace fine-grained-tool-streaming-2025-05-14 with eager_input_streaming
• Plan for Amazon Bedrock and Google-managed Vertex AI having their own retirement dates

Last updated 2026-11-02.

Check your code before Nov 30

Related tools