Claude Sonnet 4.5

A Dated Snapshot From the Manual-Reasoning Era

Anthropic launched Claude Sonnet 4.5 on September 29, 2025 and described it at launch as its best model for complex agents and coding.

The pinned Claude API model ID is claude-sonnet-4-5-20250929, while claude-sonnet-4-5 is the provider alias.

The model provides a 200,000-token context window and supports up to 64,000 output tokens. It accepts text and images and produces text.

Its reasoning system is fundamentally different from Sonnet 4.6 and newer generations. Sonnet 4.5 supports manual extended thinking only. Applications explicitly enable thinking and allocate a budget_tokens value rather than using adaptive thinking and effort.

Anthropic deprecated Sonnet 4.5 on September 30, 2026. It remains functional during the deprecation window, but Claude API retirement is scheduled for November 30, 2026. Anthropic recommends Claude Sonnet 5.5 as the replacement.

• Model ID: claude-sonnet-4-5-20250929.
• Released September 29, 2025.
• Deprecated September 30, 2026.
• Claude API retirement scheduled for November 30, 2026.
• 200K context window.
• 64K maximum output.
• Text and image input with text output.
• Manual extended thinking only.

Reasoning Depth Is an Explicit Token Budget

Thinking is enabled with thinking: {"type": "enabled", "budget_tokens": N}.

The minimum budget_tokens value is 1,024 tokens.

Outside interleaved-thinking mode, the thinking budget must be smaller than max_tokens, because reasoning tokens and the visible final response share the same output ceiling.

The budget is a target rather than a promise that Claude will consume every allocated token. The model can finish reasoning early when the problem does not require the full budget.

This gives a legacy application a direct control that newer adaptive-thinking models intentionally remove: a numerical reasoning allowance.

That can be useful when a workload was tuned around predictable maximum reasoning spend or latency. It also creates migration work, because Sonnet 5.5 does not accept manual budget_tokens.

Sonnet 4.5 does not support adaptive thinking. Sending thinking.type: "adaptive" returns an error.

Legacy Economics Before the $2/$10 Sonnet Generation

Anthropic lists 5-minute cache writes at $3.75 per million tokens, 1-hour cache writes at $6 per million, and cache reads at $0.30 per million.

Prompt caching can reduce repeated input cost when the application reuses system instructions, repository content, policies, documents, tool definitions, or other stable prefixes.

Batch API processing receives a 50% discount on standard input and output pricing.

Current Sonnet 5.5 is cheaper at $2 input and $10 output per million tokens. The nominal rate reduction is significant, but migration economics are not a direct price-ratio calculation because newer Claude generations also use different tokenization.

The right comparison is cost per accepted result across the same workload, including thinking, tools, retries, cache operations, and any fallback calls.

• Input: $3.00 per 1M tokens.
• Output: $15.00 per 1M tokens.
• 5-minute cache write: $3.75 per 1M tokens.
• 1-hour cache write: $6.00 per 1M tokens.
• Cache read: $0.30 per 1M tokens.
• Batch API: 50% discount on input and output.

Tool Use Has Manual-Thinking Constraints

Interleaved thinking lets Claude examine a tool result and reason again before choosing the next action.

On Sonnet 4.5, this behavior is not automatic. The application enables it through the legacy beta header.

Manual extended thinking also restricts tool_choice. While thinking is enabled, Anthropic supports automatic tool selection or no tool. Forced tool_choice: any and forced named-tool selection are incompatible with manual extended thinking.

Response prefilling follows a similar pattern: older models can use prefilled assistant responses when thinking is off, but prefilling cannot be combined with active thinking.

Thinking-block persistence is another generational difference. Sonnet 4.5 strips earlier thinking blocks from context when later non-tool-result user content is added. Sonnet 4.6 and newer preservation behavior is different.

That can affect prompt caching and long conversation replay, so migrate complete multi-turn tool traces rather than evaluating only isolated requests.

This Is Not a Model-ID-Only Upgrade

Sonnet 4.5 runs without thinking unless manual extended thinking is explicitly enabled.

Sonnet 5.5 runs with adaptive thinking when the thinking field is omitted. It does not accept thinking.type: "enabled" or budget_tokens.

That means the old reasoning configuration must be removed rather than copied.

Capacity changes from 200K context / 64K output to 1M context / 128K standard output.

Standard pricing falls from $3/$15 to $2/$10 per million input/output tokens.

Tool behavior also changes. Manual extended thinking’s legacy orchestration patterns do not map directly to the current model, and Sonnet 5.5 itself does not support forced any-tool or named-tool choice.

Assistant prefill should also be removed. Claude 4.6 and later models reject prefilled assistant messages entirely.

Finally, re-evaluate token counts and retained conversation state. The tokenizer generation changes, and newer models preserve thinking state differently.