Claude API Request Parameters Cheatsheet

A dense reference for the parameters you'll pass to client.messages.create(...) on nearly every call, plus the ones you'll reach for less often.

Use it to look up a parameter's type, default, and when to adjust it, without digging through prose docs.

How to Use This Cheatsheet

Start with model and max_tokens, the only two required parameters.
Adjust temperature or top_p when output feels too repetitive or too erratic, but usually only one of the two.
Treat top_k as a rarely-needed, more aggressive sibling of top_p.
Revisit this table whenever you're tuning a new use case's determinism or output length.

Core Parameters

Parameter	Type	Required?	Default	Description
`model`	`string`	Yes	none	The model ID to use, e.g. `claude-sonnet-5-20260630`. Exact ID strings change over time - verify current values in the docs before hardcoding.
`max_tokens`	`integer`	Yes	none	Hard ceiling on tokens Claude may generate in the response. Generation stops immediately when reached, even mid-sentence.
`messages`	`array`	Yes	none	Ordered list of `{"role": "user"
`system`	`string`	No	none	Persistent instructions applied across the whole conversation, set outside the `messages` array.
`temperature`	`number`	No	`1.0`	Controls randomness, `0` is near-deterministic, higher values are more varied. Range `0` to `1`.
`top_p`	`number`	No	`1.0`	Nucleus sampling - restricts token choice to the smallest set whose cumulative probability reaches `top_p`.
`top_k`	`integer`	No	none (unset)	Restricts token choice to the `k` most likely next tokens. Coarser and less commonly used than `top_p`.
`stream`	`boolean`	No	`false`	When `true`, the response is returned as a sequence of server-sent events instead of one JSON object.
`stop_sequences`	`array` of `string`	No	none	Custom strings that, if generated, stop the response early.

Model ID Reference

Model	Illustrative ID pattern	Tier
Claude Fable 5	`claude-fable-5-20260615`	Top / Mythos-class, 1M context, 128K max output
Claude Opus 4.8	`claude-opus-4-8-20260415`	Flagship reasoning, 1M context by default
Claude Sonnet 5	`claude-sonnet-5-20260630`	Default model, released 2026-06-30
Claude Haiku 4.5	`claude-haiku-4-5-20260601`	Fastest / cheapest, 200K context

Model ID strings shown here follow the current naming pattern (claude-<name>-<version>-<date>) but should be verified against the live docs before use, exact strings and availability change as new versions ship.

temperature vs. top_p vs. top_k

Parameter	What It Controls	Typical Range to Adjust	Combine With
`temperature`	Overall randomness of token selection	`0.0`-`0.3` for factual/structured tasks, `0.7`-`1.0` for creative tasks	Usually adjusted alone
`top_p`	Nucleus size, how much of the probability mass is eligible	`0.9`-`1.0` for most tasks	Often left at default while tuning `temperature`
`top_k`	Hard cutoff on candidate token count	Rarely changed from unset; `40`-`100` if used	Rarely combined with aggressive `temperature` changes

Usage Examples

import anthropic
 
client = anthropic.Anthropic()
 
# Deterministic, factual task: low temperature
factual = client.messages.create(
    model="claude-sonnet-5-20260630",
    max_tokens=100,
    temperature=0.0,
    messages=[{"role": "user", "content": "What is the boiling point of water at sea level?"}],
)
 
# Creative brainstorming: higher temperature
creative = client.messages.create(
    model="claude-sonnet-5-20260630",
    max_tokens=300,
    temperature=0.9,
    messages=[{"role": "user", "content": "Brainstorm five taglines for a coffee shop."}],
)
 
# Streaming a longer response
with client.messages.stream(
    model="claude-opus-4-8-20260415",
    max_tokens=1000,
    messages=[{"role": "user", "content": "Write a short story about a lighthouse."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Frequently Paired Parameters

Parameter	Type	Pairs Well With	Why
`system`	`string`	`temperature`	A focused system prompt plus low `temperature` gives consistent, on-task output.
`stream`	`boolean`	`max_tokens`	Streaming with a generous `max_tokens` suits long-form generation shown incrementally.
`stop_sequences`	`array`	`max_tokens`	Combine a hard token cap with a semantic stop point for structured formats like JSON.

FAQs

Which parameters are required on every call?

model, max_tokens, and messages. Everything else is optional and has a documented default.

What happens if I don't set temperature?

It defaults to 1.0, the most varied setting; for tasks needing consistent, repeatable output, set it lower explicitly.

Should I adjust both temperature and top_p at the same time?

Generally no. Pick one to tune for a given task, adjusting both aggressively at once makes behavior harder to predict and debug.

What's the difference between top_p and top_k?

top_p restricts the candidate pool by cumulative probability (a soft, adaptive cutoff).
top_k restricts it to a fixed number of top candidates (a hard, static cutoff).

Does max_tokens control how long the response will be?

It sets an upper limit, not a target; Claude may stop naturally well before reaching it, or get cut off if it hits the ceiling mid-response.

Where do system prompts go in the request?

In a top-level system parameter, separate from the messages array, not as a message with a "system" role.

Are the model ID strings shown here guaranteed to be correct?

No, they illustrate the current naming pattern as of ~June 2026 but exact strings and available dates change; always verify against the live API docs before hardcoding a model ID in production.

What does stop_sequences do that max_tokens doesn't?

max_tokens is a hard length ceiling regardless of content; stop_sequences stops generation as soon as a specific string appears, useful for structured output formats.

Is stream a required parameter?

No, it defaults to false; set it to true only when you want incrementally delivered output instead of one final JSON response.

Which parameter should I lower first if output feels too random?

temperature, it's the most direct and commonly used control; only reach for top_p or top_k if lowering temperature alone doesn't get the behavior you want.

Understanding the Claude API Request Lifecycle - where these parameters fit in a request.
Claude API Fundamentals Basics - these parameters used in runnable examples.
Selecting a Model: Claude Fable 5, Opus 4.8, Sonnet 5, and Haiku 4.5 - choosing the right model value.

Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5 - and the official anthropic Python SDK (latest 0.x release). Model names, SDK versions, and pricing move quickly - verify current specifics at platform.claude.com/docs before relying on them.

Claude API Request Parameters Cheatsheet

A dense reference for the parameters you'll pass to client.messages.create(...) on nearly every call, plus the ones you'll reach for less often.

Use it to look up a parameter's type, default, and when to adjust it, without digging through prose docs.

How to Use This Cheatsheet

Start with model and max_tokens, the only two required parameters.
Adjust temperature or top_p when output feels too repetitive or too erratic, but usually only one of the two.
Treat top_k as a rarely-needed, more aggressive sibling of top_p.
Revisit this table whenever you're tuning a new use case's determinism or output length.

Core Parameters

Parameter	Type	Required?	Default	Description
`model`	`string`	Yes	none	The model ID to use, e.g. `claude-sonnet-5-20260630`. Exact ID strings change over time - verify current values in the docs before hardcoding.
`max_tokens`	`integer`	Yes	none	Hard ceiling on tokens Claude may generate in the response. Generation stops immediately when reached, even mid-sentence.
`messages`	`array`	Yes	none	Ordered list of `{"role": "user"
`system`	`string`	No	none	Persistent instructions applied across the whole conversation, set outside the `messages` array.
`temperature`	`number`	No	`1.0`	Controls randomness, `0` is near-deterministic, higher values are more varied. Range `0` to `1`.
`top_p`	`number`	No	`1.0`	Nucleus sampling - restricts token choice to the smallest set whose cumulative probability reaches `top_p`.
`top_k`	`integer`	No	none (unset)	Restricts token choice to the `k` most likely next tokens. Coarser and less commonly used than `top_p`.
`stream`	`boolean`	No	`false`	When `true`, the response is returned as a sequence of server-sent events instead of one JSON object.
`stop_sequences`	`array` of `string`	No	none	Custom strings that, if generated, stop the response early.

Model ID Reference

Model	Illustrative ID pattern	Tier
Claude Fable 5	`claude-fable-5-20260615`	Top / Mythos-class, 1M context, 128K max output
Claude Opus 4.8	`claude-opus-4-8-20260415`	Flagship reasoning, 1M context by default
Claude Sonnet 5	`claude-sonnet-5-20260630`	Default model, released 2026-06-30
Claude Haiku 4.5	`claude-haiku-4-5-20260601`	Fastest / cheapest, 200K context

temperature vs. top_p vs. top_k

Parameter	What It Controls	Typical Range to Adjust	Combine With
`temperature`	Overall randomness of token selection	`0.0`-`0.3` for factual/structured tasks, `0.7`-`1.0` for creative tasks	Usually adjusted alone
`top_p`	Nucleus size, how much of the probability mass is eligible	`0.9`-`1.0` for most tasks	Often left at default while tuning `temperature`
`top_k`	Hard cutoff on candidate token count	Rarely changed from unset; `40`-`100` if used	Rarely combined with aggressive `temperature` changes

Usage Examples

import anthropic
 
client = anthropic.Anthropic()
 
# Deterministic, factual task: low temperature
factual = client.messages.create(
    model="claude-sonnet-5-20260630",
    max_tokens=100,
    temperature=0.0,
    messages=[{"role": "user", "content": "What is the boiling point of water at sea level?"}],
)
 
# Creative brainstorming: higher temperature
creative = client.messages.create(
    model="claude-sonnet-5-20260630",
    max_tokens=300,
    temperature=0.9,
    messages=[{"role": "user", "content": "Brainstorm five taglines for a coffee shop."}],
)
 
# Streaming a longer response
with client.messages.stream(
    model="claude-opus-4-8-20260415",
    max_tokens=1000,
    messages=[{"role": "user", "content": "Write a short story about a lighthouse."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Frequently Paired Parameters

Parameter	Type	Pairs Well With	Why
`system`	`string`	`temperature`	A focused system prompt plus low `temperature` gives consistent, on-task output.
`stream`	`boolean`	`max_tokens`	Streaming with a generous `max_tokens` suits long-form generation shown incrementally.
`stop_sequences`	`array`	`max_tokens`	Combine a hard token cap with a semantic stop point for structured formats like JSON.

FAQs

Which parameters are required on every call?

model, max_tokens, and messages. Everything else is optional and has a documented default.

What happens if I don't set temperature?

It defaults to 1.0, the most varied setting; for tasks needing consistent, repeatable output, set it lower explicitly.

Should I adjust both temperature and top_p at the same time?

Generally no. Pick one to tune for a given task, adjusting both aggressively at once makes behavior harder to predict and debug.

What's the difference between top_p and top_k?

top_p restricts the candidate pool by cumulative probability (a soft, adaptive cutoff).
top_k restricts it to a fixed number of top candidates (a hard, static cutoff).

Does max_tokens control how long the response will be?

It sets an upper limit, not a target; Claude may stop naturally well before reaching it, or get cut off if it hits the ceiling mid-response.

Where do system prompts go in the request?

In a top-level system parameter, separate from the messages array, not as a message with a "system" role.

Are the model ID strings shown here guaranteed to be correct?

No, they illustrate the current naming pattern as of ~June 2026 but exact strings and available dates change; always verify against the live API docs before hardcoding a model ID in production.

What does stop_sequences do that max_tokens doesn't?

max_tokens is a hard length ceiling regardless of content; stop_sequences stops generation as soon as a specific string appears, useful for structured output formats.

Is stream a required parameter?

No, it defaults to false; set it to true only when you want incrementally delivered output instead of one final JSON response.

Which parameter should I lower first if output feels too random?

temperature, it's the most direct and commonly used control; only reach for top_p or top_k if lowering temperature alone doesn't get the behavior you want.

Understanding the Claude API Request Lifecycle - where these parameters fit in a request.
Claude API Fundamentals Basics - these parameters used in runnable examples.
Selecting a Model: Claude Fable 5, Opus 4.8, Sonnet 5, and Haiku 4.5 - choosing the right model value.

Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5 - and the official anthropic Python SDK (latest 0.x release). Model names, SDK versions, and pricing move quickly - verify current specifics at platform.claude.com/docs before relying on them.

Claude API Request Parameters Cheatsheet

Search Documentation

Claude API Request Parameters Cheatsheet