Claude Haiku 4.5

Claude Haiku 4.5 is the cheapest and quickest model in the catalog — a dollar per million input tokens, with a 200K context window and a 64K output cap. It uses the older extended-thinking parameter rather than adaptive thinking, which makes it predictable to budget for. Built for the calls you make thousands of times an hour.

Input / output price: From $1 / $5 per 1M
Context: 200K tokens
Max output: 64K tokens

index.ts

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.onefoldhq.com/v1",
  apiKey: process.env.ONEFOLD_API_KEY,
});

const result = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [{ role: "user", content: "Why is the sky blue?" }],
});

console.log(result.choices[0].message.content);

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.onefoldhq.com/v1",
  apiKey: process.env.ONEFOLD_API_KEY,
});

const result = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [{ role: "user", content: "Why is the sky blue?" }],
});

console.log(result.choices[0].message.content);

main.py

from openai import OpenAI

client = OpenAI(
    base_url="https://api.onefoldhq.com/v1",
    api_key=os.environ["ONEFOLD_API_KEY"],
)

result = client.chat.completions.create(
    model="anthropic/claude-haiku-4-5",
    messages=[{"role": "user", "content": "Why is the sky blue?"}],
)

print(result.choices[0].message.content)

from openai import OpenAI

client = OpenAI(
    base_url="https://api.onefoldhq.com/v1",
    api_key=os.environ["ONEFOLD_API_KEY"],
)

result = client.chat.completions.create(
    model="anthropic/claude-haiku-4-5",
    messages=[{"role": "user", "content": "Why is the sky blue?"}],
)

print(result.choices[0].message.content)

request.sh

curl https://api.onefoldhq.com/v1/chat/completions \
  -H "Authorization: Bearer $ONEFOLD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-haiku-4-5",
    "messages": [{ "role": "user", "content": "Why is the sky blue?" }]
  }'

curl https://api.onefoldhq.com/v1/chat/completions \
  -H "Authorization: Bearer $ONEFOLD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-haiku-4-5",
    "messages": [{ "role": "user", "content": "Why is the sky blue?" }]
  }'

Playground

Try out Claude Haiku 4.5 by Anthropic. Usage is billed to your team at API rates.

Claude Haiku 4.5

Message Claude Haiku 4.5

Providers

Route to any available provider. One request shape for all of them, and the router picks on price, health and latency.

Provider	Context	Max output	Latency	Throughput	Input	Output	Cache
AnthropicLegal: Terms · Privacy	200K	64K	0.6 s	49 tps	$1/M	$5/M	Read$0.10/MWrite$1.25/M
BedrockLegal: Terms · Privacy	200K	8K	1.3 s	63 tps	$1/M+3 more	$5/M+3 more	Read$0.10/MWrite$1.25/M
Google Vertex AILegal: Terms · Privacy	200K	8K	1.1 s	41 tps	$1/M+3 more	$5/M+3 more	Read$0.10/MWrite$1.25/M

Uptime 24 hours

Direct request success rate on Onefold and per-provider. Visit the docs for more info.

Sep 10, 8:45 PMSep 11, 8:30 PM

Uptime
Series	Low	Typical	High
Onefold	100.00%	100.00%	100.00%
Anthropic	99.59%	99.95%	99.98%
Bedrock	97.21%	99.88%	99.99%
Google Vertex AI	87.29%	99.20%	99.97%

1Onefold100.00%
2Anthropic99.95%
3Bedrock99.88%
4Google Vertex AI99.20%

Status 24 hours

Reliability of providers. Visit the docs for more info.

Onefold100.00%
Anthropic99.95%
Bedrock99.88%
Google Vertex AI99.20%

Sep 10, 8:45 PMSep 11, 8:30 PM

Status of Onefold and its providers over the last 24 hours
Provider	Uptime	Operational windows	Degraded windows	Outage windows
Onefold	100.00%	96	0	0
Anthropic	99.95%	96	0	0
Bedrock	99.88%	92	4	0
Google Vertex AI	99.20%	85	6	5

Throughput 24 hours

P50 throughput on live gateway traffic, in tokens per second (TPS). Visit the docs for more info.

Sep 10, 8:45 PMSep 11, 8:30 PM

Throughput
Series	Low	Typical	High
Anthropic	46.05 tps	47.48 tps	52.95 tps
Bedrock	65.26 tps	66.81 tps	69.34 tps
Google Vertex AI	57.34 tps	59.04 tps	60.19 tps

1Bedrock66.81 tps
2Google Vertex AI59.04 tps
3Anthropic47.48 tps

Latency 24 hours

P50 time to first token (TTFT) on live gateway traffic, in seconds. View the docs for more info.

Sep 10, 8:45 PMSep 11, 8:30 PM

Latency
Series	Low	Typical	High
Anthropic	0.74 s	0.83 s	0.97 s
Bedrock	1.31 s	1.55 s	1.60 s
Google Vertex AI	1.04 s	1.19 s	1.26 s

1Anthropic0.83 s
2Google Vertex AI1.19 s
3Bedrock1.55 s

Pricing with caching

What input tokens cost as more of your prefix is reused. Cached input bills at $0.10 per million instead of $1.

Chart requires JavaScript. The figures are in the table below.

Blended input rate
List price

View as table

Blended input price for Claude Haiku 4.5 by cache hit ratio
Cache hit ratio	0%	25%	50%	75%	90%
Blended input rate	$1	$0.78	$0.55	$0.33	$0.19
List price	$1	$1	$1	$1	$1

Cache hit ratio	Blended input	Saving	1M in + 1M out
0%	$1/M	—	$6
25%	$0.78/M	22%	$5.78
50%	$0.55/M	45%	$5.55
75%	$0.33/M	68%	$5.33
90%	$0.19/M	81%	$5.19

Getting started

Call Claude Haiku 4.5 through Onefold with any OpenAI-compatible client by changing the base URL. Onefold authenticates the request and routes it to an available provider.

index.ts

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.onefoldhq.com/v1",
  apiKey: process.env.ONEFOLD_API_KEY,
});

const result = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [{ role: "user", content: "Why is the sky blue?" }],
});

console.log(result.choices[0].message.content);

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.onefoldhq.com/v1",
  apiKey: process.env.ONEFOLD_API_KEY,
});

const result = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [{ role: "user", content: "Why is the sky blue?" }],
});

console.log(result.choices[0].message.content);

Streaming

Stream tokens as they are produced. The gateway never buffers the upstream response.

stream.ts

const stream = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [{ role: "user", content: "Summarise this document." }],
  stream: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

const stream = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [{ role: "user", content: "Summarise this document." }],
  stream: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

Image input

Send images alongside text as message parts. Images count as input tokens.

image-input.ts

const result = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [
    {
      role: "user",
      content: [
        { type: "text", text: "Describe this image." },
        { type: "image_url", image_url: { url: "https://example.com/photo.jpg" } },
      ],
    },
  ],
});

const result = await client.chat.completions.create({
  model: "anthropic/claude-haiku-4-5",
  messages: [
    {
      role: "user",
      content: [
        { type: "text", text: "Describe this image." },
        { type: "image_url", image_url: { url: "https://example.com/photo.jpg" } },
      ],
    },
  ],
});

Top-level parameters

The same Claude Haiku 4.5 request in every API format Onefold supports.

Parameter	Type	Required	Description
model	string	Yes	Model id in the form `creator/model`, here `anthropic/claude-haiku-4-5`. Onefold routes the request to an available provider.
messages	Message[]	Yes	The conversation so far. System, user, assistant and tool roles are all accepted and translated into the provider dialect.
max_tokens	number	No	Hard cap on generated tokens. Claude Haiku 4.5 supports up to `64,000`. Thinking tokens count toward this limit.
stream	boolean	No	Stream the response as server-sent events. The gateway streams byte by byte rather than buffering the upstream response.
temperature	number	No	Sampling temperature. Passed through to the provider unchanged.
tools	Tool[]	No	Tool definitions in OpenAI shape. Translated into the provider native format, including parallel tool calls.

Input limits

Input	Formats	Sources	Limits
Text	—	—	Prompt and response share the 200K-token context window
Image	png, jpeg, webp, gif	URL, base64	Sent as image parts in messages; counts as input tokens
PDF	pdf	URL, base64	Sent as file parts in messages; counts as input tokens

Capabilities

Streaming
Tool use
Vision
PDF input
Citations
Structured outputs
Extended thinking
Prompt caching
Batch API

Accepted parameters

max_tokens
stream
stop_sequences
system
tools
tool_choice
thinking
temperature
top_p
top_k
output_config.format

Specification

Model id: anthropic/claude-haiku-4-5
Upstream id: claude-haiku-4-5
Modalities: Text, Image in · Text out
Context window: 200,000 tokens
Max output: 64,000 tokens
API dialect: anthropic, translated for you
Knowledge cutoff: Feb 2025
Min cacheable prefix: 4,096 tokens
Available on: Anthropic, Bedrock, Google Vertex AI

Similar models

Switch between any of them by changing the model string.

Model	Context	Input	Output
anthropic/claude-sonnet-5	1M	$3/M	$15/M
anthropic/claude-sonnet-4-6	1M	$3/M	$15/M
anthropic/claude-sonnet-4-5	200K	$3/M	$15/M

About Claude Haiku 4.5

It speaks the anthropic dialect and reaches 3 platforms through Onefold. Its knowledge is most reliable through Feb 2025. Anthropic released it on 1 Oct 2025.

What to consider when choosing a provider

Output ceiling: platforms cap output independently of the model. Anthropic allows 64K. Bedrock allows 8K. Google Vertex AI allows 8K.
Authentication: Onefold authenticates requests with one key. You do not manage provider credentials directly.
Dialect: this model speaks anthropic. You never write it — the gateway accepts OpenAI-shaped requests and translates at the edge.

When to use Claude Haiku 4.5

Best for

Classification, extraction, routing and other high-frequency short work.
Agentic tool use: parallel tool calls, translated to the native format

Consider alternatives when

You need more context: Claude Sonnet 5 carries 1M
Latency dominates: a lighter model in the same family answers sooner
The task is narrow: a smaller model hits the same bar for less on well-scoped work

Frequently asked questions

What does Claude Haiku 4.5 cost?

$1 per million input tokens and $5 per million output tokens, which is Anthropic's published list price. Onefold adds a flat 5% routing fee on top, billed from prepaid credits. Cached input is far cheaper at $0.10 per million tokens.

What is the context window for Claude Haiku 4.5?

200,000 tokens of context, with up to 64,000 tokens in a single response. Its knowledge is most reliable through Feb 2025.

How do I call Claude Haiku 4.5?

Point any OpenAI-compatible client at https://api.onefoldhq.com/v1 and pass "anthropic/claude-haiku-4-5" as the model string. There is no SDK to install, we translate the request into Anthropic's anthropic dialect on the way out and the response stream back on the way in.

How is prompt caching billed?

Reading from cache costs $0.10 per million tokens, one tenth of the list input price. Writing to a five-minute cache costs $1.25 (1.25× list), and a one-hour cache costs $2 (2× list). Prefixes shorter than 4,096 tokens will not cache at all.

Which API dialect does Claude Haiku 4.5 speak?

The anthropic dialect. You never write it: the gateway accepts OpenAI-shaped requests and translates them at the edge, streaming byte by byte rather than buffering. Switching to a model on a different dialect means changing the model string and nothing else.

Is Claude Haiku 4.5 being retired?

No retirement date has been announced. Anthropic publishes deprecations ahead of time, and this page tracks that field, so a date will appear here if one is set.