Claude Haiku 4.5
Claude Haiku 4.5 by Anthropic: $1 per 1M input and $5 per 1M output tokens, 200K context. Call it with one OpenAI-compatible API key, compare it with similar models, start free.
Claude Haiku 4.5 is served by Anthropic and called through odnoga on /v1/chat/completions, in the same OpenAI-compatible shape every other model on that endpoint uses. It is a reasoning model, so it spends extra output tokens working through a problem before answering — budget for higher output cost on hard tasks. It accepts images alongside text. It supports tool and function calling. It can be forced to return structured JSON. Its 200K-token context window sets how much input you can send in one request. Cached input is billed at $0.1 per 1M tokens, so repeated prefixes cost less.
Specification and price
| Vendor | Anthropic |
|---|---|
| Model ID | claude-haiku-4-5 |
| Context window | 200K tokens |
| Max output | 64K tokens |
| Served on | /v1/chat/completions |
| Capabilities | Reasoning, Reads images, Tool calling, JSON schema, Streaming, Web search, Prompt caching, Batch |
| Per 1M tokens | Vendor cost | Your price on Free+7% |
|---|---|---|
| Input | $1 | $1.08 |
| Output | $5 | $5.38 |
| Cached input | $0.1 | $0.108 |
Vendor cost is the list price per million tokens as recorded in the odnoga catalog; your price applies your plan margin with the same formula that bills every request — see pricing. Pricing
Call it through odnoga
const res = await fetch("https://api.odnoga.com/v1/chat/completions", {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.ODNOGA_API_KEY}`,
"Content-Type": "application/json",
},
body: JSON.stringify({
model: "claude-haiku-4-5",
messages: [{ role: "user", content: "Hello" }],
}),
});
One endpoint, one key, one bill — /v1/chat/completions takes the same shape whichever vendor is behind it. Swap the model id and odnoga handles keys, routing, limits and cost accounting.
What Claude Haiku 4.5 can do
Everything below is what the model itself supports. Ask for something it does not do and odnoga tells you before the request reaches the provider, instead of returning a worse answer.
Works through a problem step by step before answering, which costs more tokens and buys accuracy on hard tasks.
Send a screenshot, a photo or a scan in the same message as your text.
The model can call functions you define and use what they return.
Answers with valid JSON instead of prose you would have to parse.
Answers in the exact shape you specify, so the response fits your types without a fallback path.
Tokens arrive as they are produced, so your interface can answer immediately.
Can look something up rather than answer from training data alone. Billed per search.
A prompt you send repeatedly is charged at the cached rate, which is a fraction of the input price.
Work that can wait goes through at a lower price.
Finer control
- Structured outputs · The provider guarantees the schema rather than merely aiming for it.
- Parallel tools · Can call several of your functions in one turn instead of one after another.
How to use Claude Haiku 4.5
Reviewed 2026-09-14
Best for
- High-frequency work that still needs good judgement: Anthropic calls it the fastest model with near-frontier intelligence.
- Interactive features where response time is part of the product.
- Classification, extraction, summarisation and routing at scale.
- A cheap first stage in front of Sonnet 5 or Opus 5.
Not the right pick when
- Long autonomous agent runs and deep multi-step reasoning.
- Very large inputs: the window is 200K tokens, not the 1M of the Sonnet and Opus 5 generation.
- Final customer-facing copy where nuance decides the outcome.
Practical tips through odnoga
- 01Make the task closed and the output schema explicit — that is where fast models are most reliable.
- 02Route by difficulty: Haiku first, escalate on low confidence. odnoga records both legs so you can prove the split is cheaper.
- 03Enable odnoga answer reuse for deterministic calls; identical repeats then cost nothing at the vendor.
- 04Check the 200K context limit before reusing a prompt written for a 1M-token model.
What a month costs
1,000 calls a month, 10,000 input tokens and 2,000 output tokens each, at your Free price (vendor cost +7%):
| Input (10M tokens) | $10.75 |
|---|---|
| Output (2M tokens) | $10.75 |
| Your cost per month on Free | $21.51 |
Vendor list cost $20.00 + odnoga margin $1.51 (+7%). Cached input or answer reuse lowers it; odnoga records both numbers per request.
If Claude Haiku 4.5 is not the one
Models that take the same kind of input, are priced the same way, and do the closest thing to what this one does. Not a quality ranking — odnoga does not benchmark models.
| Model | Context | Input / 1M | Output / 1M | Capabilities |
|---|---|---|---|---|
| Claude Haiku 4.5 | 200K | $1 | $5 | Reasoning, Reads images, Tool calling, JSON schema, Streaming, Web search, Prompt caching, Batch |
| grok-build-0.1 | 256K | $1 | $2 | Reasoning, Reads images, Tool calling, JSON schema, Streaming, Web search, Prompt caching |
| o4-mini | 200K | $1.10 | $4.40 | Reasoning, Reads images, Tool calling, JSON schema, Streaming, Prompt caching, Batch |
| o3-mini | 200K | $1.10 | $4.40 | Reasoning, Tool calling, JSON schema, Streaming, Prompt caching, Batch |
Questions
- How fast is Claude Haiku 4.5 really?
- Anthropic positions it as its fastest model with near-frontier intelligence, at 200K context and 64K max output. Measure the latency on your own prompts — the evaluation laboratory records median latency per model.
- Can Haiku replace Sonnet in production?
- For closed, well-specified tasks it often can. Freeze representative test cases, run both, and let the pass rate and cost columns decide.
- What is the cheapest safe setup?
- Haiku as the default with a stronger fallback on failure or low confidence, a per-tenant budget, and answer reuse on deterministic prompts.
All models · Pricing · Docs · Compare models in the evaluation lab
Sources: Anthropic — Claude Haiku 4.5 overview, Anthropic — models overview