Claude Sonnet 5 API Errors: Why Temperature, Top P, and Manual Thinking May Return 400

Q: Why does Claude Sonnet 5 return a 400 error for `temperature`?

Claude Sonnet 5 does not accept `temperature` set to a non-default value. Remove the parameter and use system-prompt instructions to control tone, style, and variation.

Q: Does Claude Sonnet 5 support `top_p`?

Claude Sonnet 5 does not accept `top_p` set to a non-default value. The safest migration path is to omit it from the request payload.

Q: Does Claude Sonnet 5 support `top_k`?

Claude Sonnet 5 does not accept `top_k` set to a non-default value. Remove it when migrating from older request templates.

Q: Why does manual extended thinking return a 400 error?

Manual extended thinking with `thinking: {type: "enabled", budget_tokens: N}` was deprecated on Claude Sonnet 4.6 and removed on Claude Sonnet 5. Use adaptive thinking instead.

Q: Is adaptive thinking on by default?

Yes. Claude Platform Docs state that requests without a `thinking` field run with adaptive thinking on Claude Sonnet 5. To turn thinking off, pass `thinking: {type: "disabled"}`.

Q: How do I control reasoning depth without `budget_tokens`?

Use adaptive thinking and the `effort` parameter where supported. For hard coding or agentic tasks, raise effort to `high` or `xhigh`. For simpler or lower-latency tasks, use lower effort or disable thinking where appropriate.

Q: Why is my Claude Sonnet 5 answer truncated?

Your `max_tokens` may be too low. Adaptive thinking uses part of the output budget, and Claude Sonnet 5's new tokenizer can produce more tokens for the same text. Increase `max_tokens`, lower effort, or ask for a shorter final answer.

Introduction

Claude Sonnet 5 is designed as a drop-in upgrade from Claude Sonnet 4.6, but "drop-in" does not mean every old request payload will work unchanged. Some common parameters that developers used with earlier Claude models can now return a 400 error.

The most common causes are:

Setting temperature to a non-default value.
Setting top_p to a non-default value.
Setting top_k to a non-default value.
Using manual extended thinking with thinking: {type: "enabled", budget_tokens: N}.
Prefilling the assistant message, which was already unsupported on Claude Sonnet 4.6.

Claude Platform Docs state that Claude Sonnet 5 has three important behavior changes from Claude Sonnet 4.6: adaptive thinking is on by default, manual extended thinking now returns a 400 error, and non-default sampling parameters such as temperature, top_p, and top_k return a 400 error.

This guide explains why these errors happen, how to fix them, and how to safely migrate old Claude Sonnet 4.6 request payloads to Claude Sonnet 5. It also shows how to think about these changes when using Claude Sonnet 5 through iCreat API, an OpenAI-compatible unified API platform.

Quick Fix

If your Claude Sonnet 5 request returns a 400 error, check your request body first.

Remove these parameters if they are set to non-default values

{
  "temperature": 0.2,
  "top_p": 0.9,
  "top_k": 40
}

For Claude Sonnet 5, the safest migration path is to omit these parameters entirely and use system-prompt instructions to guide style, tone, format, and consistency. Claude Platform Docs explicitly say that setting temperature, top_p, or top_k to a non-default value returns a 400 error.

Replace manual extended thinking

Do not use:

{
  "thinking": {
    "type": "enabled",
    "budget_tokens": 32000
  }
}

Use adaptive thinking instead:

{
  "thinking": {
    "type": "adaptive"
  }
}

Or omit the thinking field if adaptive thinking is already the default in your API surface. Claude Platform Docs state that manual extended thinking with thinking: {type: "enabled", budget_tokens: N} was deprecated on Claude Sonnet 4.6 and is removed on Claude Sonnet 5, where it returns a 400 error.

Why Claude Sonnet 5 Returns 400 for Old Parameters

Claude Sonnet 5 changes how developers should control generation behavior.

With older LLM APIs, developers often used sampling parameters such as temperature, top_p, and top_k to influence randomness, variety, and output style. They also used manual extended thinking controls to reserve a fixed reasoning budget.

Claude Sonnet 5 moves away from that pattern. Instead, Anthropic recommends:

Omitting unsupported sampling parameters.
Using system prompts and clear instructions for style and behavior.
Using adaptive thinking.
Controlling reasoning depth with the effort parameter where supported.
Leaving enough max_tokens headroom for thinking and final response text.

Claude's Sonnet 5 best practices guide says that if you previously relied on temperature for stylistic variety, you should remove temperature, top_p, and top_k when migrating, and use system-prompt instructions to guide tone and variety instead.

Error 1: `temperature` Returns 400

Problem

A request like this may fail:

{
  "model": "claude-sonnet-5",
  "temperature": 0.2,
  "max_tokens": 2048,
  "messages": [
    {
      "role": "user",
      "content": "Write a concise product description."
    }
  ]
}

Why it happens

Claude Sonnet 5 does not accept temperature set to a non-default value. Even if your SDK type-checks the field, the API can reject it server-side.

This matters because many OpenAI-compatible clients, internal wrappers, and shared LLM gateways include temperature by default. If your app automatically adds temperature, the migration may fail even if the rest of the payload looks correct.

Fix

Remove temperature:

{
  "model": "claude-sonnet-5",
  "max_tokens": 2048,
  "messages": [
    {
      "role": "user",
      "content": "Write a concise product description."
    }
  ]
}

Then control style with instructions:

{
  "model": "claude-sonnet-5",
  "max_tokens": 2048,
  "messages": [
    {
      "role": "system",
      "content": "Write in a concise, polished, commercial tone. Avoid exaggerated claims."
    },
    {
      "role": "user",
      "content": "Write a product description for an AI video API."
    }
  ]
}

Error 2: `top_p` Returns 400

Problem

A request like this may fail:

{
  "model": "claude-sonnet-5",
  "top_p": 0.9,
  "messages": [
    {
      "role": "user",
      "content": "Generate five alternative headlines."
    }
  ]
}

Why it happens

top_p is another sampling parameter. Claude Sonnet 5 rejects non-default top_p values with a 400 error. Claude Platform Docs recommend omitting sampling parameters entirely and using prompting to guide behavior.

Fix

Remove top_p and describe the variation you want:

{
  "model": "claude-sonnet-5",
  "messages": [
    {
      "role": "system",
      "content": "Generate varied but realistic marketing headlines. Each headline should use a different angle: speed, cost, workflow, quality, and developer experience."
    },
    {
      "role": "user",
      "content": "Generate five alternative headlines for Claude Sonnet 5 API."
    }
  ]
}

This is usually better than relying on sampling settings because it makes the desired variation explicit.

Error 3: `top_k` Returns 400

Problem

A request like this may fail:

{
  "model": "claude-sonnet-5",
  "top_k": 40,
  "messages": [
    {
      "role": "user",
      "content": "Brainstorm API use cases."
    }
  ]
}

Why it happens

top_k is also a sampling parameter. Claude Sonnet 5 does not accept non-default top_k values.

Fix

Remove top_k and prompt for diversity directly:

{
  "model": "claude-sonnet-5",
  "messages": [
    {
      "role": "system",
      "content": "Give diverse ideas across developer tools, enterprise automation, content workflows, support systems, and data analysis. Avoid repeating the same pattern."
    },
    {
      "role": "user",
      "content": "Brainstorm API use cases for Claude Sonnet 5."
    }
  ]
}

Error 4: Manual Extended Thinking Returns 400

Problem

A request like this may fail:

{
  "model": "claude-sonnet-5",
  "thinking": {
    "type": "enabled",
    "budget_tokens": 32000
  },
  "messages": [
    {
      "role": "user",
      "content": "Analyze this complex code migration plan."
    }
  ]
}

Why it happens

Manual extended thinking is removed on Claude Sonnet 5. The older pattern thinking: {type: "enabled", budget_tokens: N} returns a 400 error. Claude's Sonnet 5 prompting guide says this was deprecated on Claude Sonnet 4.6 and is now removed; developers should use adaptive thinking with the effort parameter instead.

Fix

Use adaptive thinking:

{
  "model": "claude-sonnet-5",
  "thinking": {
    "type": "adaptive"
  },
  "messages": [
    {
      "role": "user",
      "content": "Analyze this complex code migration plan."
    }
  ]
}

If your API surface supports effort, use it to control reasoning depth:

{
  "model": "claude-sonnet-5",
  "thinking": {
    "type": "adaptive"
  },
  "output_config": {
    "effort": "high"
  },
  "messages": [
    {
      "role": "user",
      "content": "Analyze this complex code migration plan."
    }
  ]
}

For very hard coding or agentic tasks, Claude's prompting guide says to raise effort to high or xhigh rather than trying to force deeper reasoning only through prompting.

Error 5: `max_tokens` Too Low Can Cause Truncated Answers

This may not return a 400 error, but it is a common migration issue.

Claude Sonnet 5 runs adaptive thinking by default. Claude Platform Docs explain that max_tokens is a hard limit on total output, including both thinking and response text, so developers should revisit max_tokens for workloads that previously ran without thinking on Claude Sonnet 4.6.

Problem

Your request succeeds, but the final answer is too short, incomplete, or truncated.

You may see something like:

stop_reason: "max_tokens"

Why it happens

Adaptive thinking can use part of the output budget. If max_tokens is too tight, the model may spend too much of the budget on reasoning and not enough on the final answer.

Claude's prompting guide warns that on long tasks, adaptive thinking can use a large share of the budget; if the budget is tight, the response may be mostly thinking followed by a truncated answer.

Fix

Increase max_tokens, lower effort, or ask for a more concise final answer.

{
  "model": "claude-sonnet-5",
  "max_tokens": 4096,
  "messages": [
    {
      "role": "system",
      "content": "Think only as much as needed. Return a concise final answer with clear recommendations."
    },
    {
      "role": "user",
      "content": "Review this architecture plan and identify the top 5 risks."
    }
  ]
}

Error 6: Assistant Message Prefilling Returns 400

This is not new to Claude Sonnet 5, but it is still worth checking.

Claude Platform Docs state that assistant message prefilling returns a 400 error on Claude Sonnet 5, unchanged from Claude Sonnet 4.6. Anthropic recommends using structured outputs, system prompt instructions, or output_config.format instead.

Problem

A request tries to prefill the assistant response:

{
  "model": "claude-sonnet-5",
  "messages": [
    {
      "role": "user",
      "content": "Return JSON."
    },
    {
      "role": "assistant",
      "content": "{"
    }
  ]
}

Fix

Use explicit output instructions instead:

{
  "model": "claude-sonnet-5",
  "messages": [
    {
      "role": "system",
      "content": "Return valid JSON only. Do not include markdown fences or explanations."
    },
    {
      "role": "user",
      "content": "Extract the product name, price, and key benefit from this text."
    }
  ]
}

If your API surface supports structured output formatting, use that instead of prefilling.

Before and After: Migrating a Sonnet 4.6 Request to Sonnet 5

Here is a typical request that may work on an older model or in an older wrapper but fail on Claude Sonnet 5:

{
  "model": "claude-sonnet-4-6",
  "temperature": 0.2,
  "top_p": 0.9,
  "top_k": 40,
  "thinking": {
    "type": "enabled",
    "budget_tokens": 16000
  },
  "max_tokens": 2048,
  "messages": [
    {
      "role": "user",
      "content": "Review this pull request and identify risky changes."
    }
  ]
}

A safer Claude Sonnet 5 version:

{
  "model": "claude-sonnet-5",
  "thinking": {
    "type": "adaptive"
  },
  "max_tokens": 4096,
  "messages": [
    {
      "role": "system",
      "content": "You are a senior software engineer. Be precise, practical, and concise. Rank issues by severity. Avoid speculation."
    },
    {
      "role": "user",
      "content": "Review this pull request and identify risky changes."
    }
  ]
}

The migration does four things:

Changes the model ID from claude-sonnet-4-6 to claude-sonnet-5.
Removes temperature, top_p, and top_k.
Replaces manual thinking with adaptive thinking.
Raises max_tokens to leave room for reasoning and final answer text.

OpenAI-Compatible API Note for iCreat Users

Claude Sonnet 5 is available on iCreat API through an OpenAI-compatible unified API. This is convenient because developers can test and route multiple models with a familiar request pattern.

However, OpenAI-compatible clients often include optional generation parameters such as temperature by habit. If your application has a shared LLM wrapper, check whether it automatically injects parameters like:

{
  "temperature": 0.7,
  "top_p": 1
}

For Claude Sonnet 5, avoid sending unsupported non-default sampling parameters. If your wrapper is used across multiple providers, add model-specific parameter filtering.

A simple rule:

If model == "claude-sonnet-5":
    remove non-default temperature
    remove non-default top_p
    remove non-default top_k
    remove manual extended thinking
    use adaptive thinking or omit thinking depending on your API surface

Check the iCreat API Docs for the latest supported fields, base URL, authentication method, and model-specific examples.

Why Prompting Replaces Temperature for Style Control

Many developers used temperature to control creativity. For Claude Sonnet 5, that approach can break the request.

Instead, use prompt-level controls.

Instead of this

{
  "temperature": 0.8,
  "messages": [
    {
      "role": "user",
      "content": "Write ad copy."
    }
  ]
}

Use this

{
  "messages": [
    {
      "role": "system",
      "content": "Write energetic but credible B2B ad copy. Use short sentences. Avoid hype. Give three distinct angles: speed, cost, and workflow."
    },
    {
      "role": "user",
      "content": "Write ad copy for a Claude Sonnet 5 API page."
    }
  ]
}

This gives you more predictable control than a sampling parameter because it defines the exact kind of variation you want.

How to Control Output Consistency Without `temperature = 0`

Some developers used temperature = 0 for determinism. The Claude migration guide notes that temperature = 0 never guaranteed identical outputs on prior models.

For more consistent output, use:

Clear system instructions.
Strict output schemas.
Few-shot examples.
Shorter task scopes.
Explicit formatting rules.
Post-processing validation.
Retry logic for invalid structured outputs.

Example:

{
  "model": "claude-sonnet-5",
  "messages": [
    {
      "role": "system",
      "content": "Return JSON only. Use exactly these keys: title, summary, risks, next_steps. Do not add extra keys."
    },
    {
      "role": "user",
      "content": "Analyze this feature brief."
    }
  ]
}

For production systems, deterministic behavior should come from schema validation and workflow design, not only from sampling settings.

How to Control Reasoning Without Manual `budget_tokens`

Manual thinking budgets are no longer supported on Claude Sonnet 5. Use adaptive thinking and effort settings where supported.

A practical approach:

Task type	Suggested approach
Simple extraction	Disable thinking or use low effort if supported
Short rewriting	Direct prompt, concise output instructions
Standard coding help	Adaptive thinking, medium or high effort
Complex code migration	Adaptive thinking, high or xhigh effort
Long agentic workflow	Adaptive thinking, high or xhigh effort, larger `max_tokens`
User-facing low-latency task	Lower effort, concise final answer

Claude's prompting guide says Sonnet 5 respects effort levels strictly, especially at the low end. If the model under-thinks at low or medium, raise effort to high or xhigh.

Troubleshooting Checklist

When Claude Sonnet 5 returns a 400 error, use this checklist.

Check	What to do
Model ID	Use `claude-sonnet-5`
`temperature`	Remove it if set to a non-default value
`top_p`	Remove it if set to a non-default value
`top_k`	Remove it if set to a non-default value
Manual thinking	Remove `thinking: {type: "enabled", budget_tokens: N}`
Adaptive thinking	Use `thinking: {type: "adaptive"}` where supported
`max_tokens`	Increase if outputs are truncated
Assistant prefill	Remove assistant message prefilling
Structured output	Use system instructions or structured output features
Shared wrapper	Filter unsupported fields per model
SDK defaults	Check whether your SDK sends parameters automatically

Common 400 Error Patterns

Pattern 1: The request worked on Sonnet 4.6 but fails on Sonnet 5

Most likely cause:

temperature, top_p, top_k, or manual thinking budget

Fix:

Remove unsupported fields and retest.

Pattern 2: The request fails even though the SDK type-checks

Most likely cause:

The SDK supports the field for older models, but Claude Sonnet 5 rejects it server-side.

Fix:

Do not rely only on SDK type definitions. Add model-specific payload filtering.

Pattern 3: The request succeeds but the answer is incomplete

Most likely cause:

max_tokens is too low for adaptive thinking plus final answer.

Fix:

Raise max_tokens, lower effort, or request a shorter final response.

Pattern 4: The model is less varied after removing temperature

Most likely cause:

The prompt does not explicitly ask for variation.

Fix:

Ask for specific types of variation in the system or user prompt.

Pattern 5: The model under-thinks complex tasks

Most likely cause:

Effort is too low or the prompt does not signal that multi-step reasoning is needed.

Fix:

Raise effort where supported, or add direct instructions that the task requires careful multi-step reasoning.

Recommended Migration Workflow

Do not migrate all requests by changing only the model name.

Use this workflow:

Find all request templates that call Claude models.
Search for temperature, top_p, top_k, and budget_tokens.
Remove unsupported sampling parameters for Claude Sonnet 5.
Replace manual extended thinking with adaptive thinking where supported.
Revisit max_tokens limits.
Recount tokens because Claude Sonnet 5 uses a new tokenizer.
Test production prompts in the iCreat Playground.
Compare success rate, cost, latency, and truncation rate.
Add model-specific payload filtering in your API wrapper.
Roll out gradually.

Claude Platform Docs also recommend recounting tokens because Claude Sonnet 5's new tokenizer produces approximately 30% more tokens for the same text, which can affect token budgets, context usage, and request cost.

Example: Model-Specific Parameter Filtering

If your application supports multiple models, add a filter before sending requests.

def sanitize_params_for_model(model: str, params: dict) -> dict:
    cleaned = dict(params)

    if model == "claude-sonnet-5":
        # Remove unsupported non-default sampling controls
        cleaned.pop("temperature", None)
        cleaned.pop("top_p", None)
        cleaned.pop("top_k", None)

        # Remove old manual extended thinking
        thinking = cleaned.get("thinking")
        if isinstance(thinking, dict) and thinking.get("type") == "enabled":
            cleaned["thinking"] = {"type": "adaptive"}

    return cleaned

This type of filtering is useful when your app routes across multiple models, such as Claude Sonnet, Claude Opus, GPT, Gemini, DeepSeek, MiniMax, and other LLMs.

FAQ

Why does Claude Sonnet 5 return a 400 error for temperature?

Claude Sonnet 5 does not accept temperature set to a non-default value. Remove the parameter and use system-prompt instructions to control tone, style, and variation.

Does Claude Sonnet 5 support top_p?

Claude Sonnet 5 does not accept top_p set to a non-default value. The safest migration path is to omit it from the request payload.

Does Claude Sonnet 5 support top_k?

Claude Sonnet 5 does not accept top_k set to a non-default value. Remove it when migrating from older request templates.

Why does manual extended thinking return a 400 error?

Manual extended thinking with thinking: {type: "enabled", budget_tokens: N} was deprecated on Claude Sonnet 4.6 and removed on Claude Sonnet 5. Use adaptive thinking instead.

Is adaptive thinking on by default?

Yes. Claude Platform Docs state that requests without a thinking field run with adaptive thinking on Claude Sonnet 5. To turn thinking off, pass thinking: {type: "disabled"}.

How do I control creativity without temperature?

Use prompt instructions. Tell the model what kind of variation you want, how many options to produce, what tone to use, and what constraints to follow.

How do I control reasoning depth without budget_tokens?

Use adaptive thinking and the effort parameter where supported. For hard coding or agentic tasks, raise effort to high or xhigh. For simpler or lower-latency tasks, use lower effort or disable thinking where appropriate.

Why is my Claude Sonnet 5 answer truncated?

Your max_tokens may be too low. Adaptive thinking uses part of the output budget, and Claude Sonnet 5's new tokenizer can produce more tokens for the same text. Increase max_tokens, lower effort, or ask for a shorter final answer.

Is a refusal the same as a 400 error?

No. Claude Platform Docs state that cybersecurity-related refusals return as a successful HTTP 200 response with stop_reason: "refusal", not as an error.

Can I use Claude Sonnet 5 on iCreat API?

Yes. Claude Sonnet 5 is available on iCreat API through an OpenAI-compatible unified API. You can test it on the Claude Sonnet 5 model page and check the iCreat API Docs for the latest request format.

Conclusion

Claude Sonnet 5 can be a strong upgrade from Claude Sonnet 4.6, but some old request patterns can now fail with a 400 error.

The most important fixes are simple:

Remove non-default temperature, top_p, and top_k.

Replace manual extended thinking with adaptive thinking.

Revisit max_tokens.

Remove assistant message prefilling.

Use prompts and structured output controls instead of unsupported generation parameters.

For production teams, the best solution is not to patch each error one by one. It is to build model-specific request sanitization, test real prompts before migration, and route each task to the right model.

With iCreat API, developers can test Claude Sonnet 5 in the Playground, integrate it through an OpenAI-compatible unified API, and compare it with other LLM, image, video, audio, 3D, and avatar models in one workflow.

Claude Sonnet 5 API Errors: Why Temperature, Top P, and Manual Thinking May Return 400

Introduction

Quick Fix

Remove these parameters if they are set to non-default values

Replace manual extended thinking

Why Claude Sonnet 5 Returns 400 for Old Parameters

Error 1: temperature Returns 400

Problem

Why it happens

Fix

Error 2: top_p Returns 400

Problem

Why it happens

Fix

Error 3: top_k Returns 400

Problem

Why it happens

Fix

Error 4: Manual Extended Thinking Returns 400

Problem

Why it happens

Fix

Error 5: max_tokens Too Low Can Cause Truncated Answers

Problem

Why it happens

Fix

Error 6: Assistant Message Prefilling Returns 400

Problem

Fix

Before and After: Migrating a Sonnet 4.6 Request to Sonnet 5

OpenAI-Compatible API Note for iCreat Users

Why Prompting Replaces Temperature for Style Control

Instead of this

Use this

How to Control Output Consistency Without temperature = 0

How to Control Reasoning Without Manual budget_tokens

Troubleshooting Checklist

Common 400 Error Patterns

Pattern 1: The request worked on Sonnet 4.6 but fails on Sonnet 5

Pattern 2: The request fails even though the SDK type-checks

Pattern 3: The request succeeds but the answer is incomplete

Pattern 4: The model is less varied after removing temperature

Pattern 5: The model under-thinks complex tasks

Recommended Migration Workflow

Example: Model-Specific Parameter Filtering

FAQ

Conclusion

Related Articles

Error 1: `temperature` Returns 400

Error 2: `top_p` Returns 400

Error 3: `top_k` Returns 400

Error 5: `max_tokens` Too Low Can Cause Truncated Answers

How to Control Output Consistency Without `temperature = 0`

How to Control Reasoning Without Manual `budget_tokens`