Claude Haiku API pricing: Anthropic, September 2026
Haiku is the cheapest Claude line on every token rate Anthropic publishes. Haiku 4.5 costs $1 input and $5 output per million tokens, exactly half of Sonnet 5 and one third of Sonnet 4.6.
Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens on Anthropic’s API. Claude Haiku 3.5 is still in Anthropic’s model price table at $0.80 input and $4 output, the lowest input price of the 18 Claude models listed. All prices are in US dollars, from Anthropic’s published pricing, checked 28 September 2026.1
Every Sonnet model costs at least twice as much per token. Haiku stays the cheapest Claude option per task unless a larger model finishes the same job on half the tokens or fewer.
TL;DR
- Haiku 4.5 charges $1 input and $5 output per million tokens. A cache hit costs $0.10, a 5-minute cache write $1.25 and a 1-hour cache write $2.1
- The Batch API halves the token rates. Haiku 4.5 drops to $0.50 input and $2.50 output, and Haiku 3.5 to $0.40 input and $2 output.1
- Haiku 3.5 is priced in Anthropic’s pricing documentation and has no card on the main pricing page. CloudPrice records a Haiku 3.5 deprecation date of 4 July 2026.1,2,3
- A request with 2,000 input tokens and 500 output tokens costs $0.0045 on Haiku 4.5 and $0.0090 on Sonnet 5 at standard rates. Sonnet 4.6 charges $0.0135 for the same token counts.
- Sonnet 5 uses a newer tokenizer that counts the same text as roughly 30% more tokens. On matching text its premium over Haiku 4.5 grows to about 2.6 times.
- Anthropic’s 1.1x US-only inference multiplier covers Claude 4.6 and later models. Haiku 4.5 and Haiku 3.5 carry earlier version numbers and bill at standard rates.1
How this ranking was built
Every Glamdring ranking follows the published research methodology.
The ranked field is the Input price under “Base tokens” in the Model pricing table of Anthropic’s pricing documentation. The unit is US dollars per million tokens (MTok), and the documentation states that all prices are in USD. The page was captured on 28 September 2026.1
The table holds 4 headline models and 14 under “Additional models”. All 18 rows are ranked. Tied models share a rank. Output costs five times input on every row, so sorting by output gives the same order.1
Anthropic’s main pricing page shows 12 model cards, 4 under “Latest models” and 8 under “Legacy models”.2
Each card’s input, output and cache prices match the documentation table. Six documentation rows have no card: Haiku 3.5, Sonnet 4, Opus 4.1, Opus 4, Mythos 5.1 and Mythos 5. The “Main pricing page” column records where each model appears.
Fast mode, cloud marketplaces and subscription plans fall outside the ranking. Anthropic states that partner-operated platforms (Bedrock and Google Cloud) have independent regional pricing.1
The documentation gives no tax treatment for direct API billing.
Claude API models ranked by input price
Claude Haiku 3.5 ranks first at $0.80 per million input tokens. Claude Haiku 4.5 ranks second at $1, and Claude Sonnet 5 third at $2.1
| Rank | Model | Main pricing page | Input ($/MTok) | Output ($/MTok) | Cache hit ($/MTok) | 5-minute cache write ($/MTok) |
|---|---|---|---|---|---|---|
| 1 | Claude Haiku 3.5 | No card | 0.80 | 4 | 0.08 | 1 |
| 2 | Claude Haiku 4.5 | Latest models | 1 | 5 | 0.10 | 1.25 |
| 3 | Claude Sonnet 5 | Latest models | 2 | 10 | 0.20 | 2.50 |
| 4= | Claude Sonnet 4.6 | Legacy models | 3 | 15 | 0.30 | 3.75 |
| 4= | Claude Sonnet 4.5 | Legacy models | 3 | 15 | 0.30 | 3.75 |
| 4= | Claude Sonnet 4 | No card | 3 | 15 | 0.30 | 3.75 |
| 7 | Claude Opus 5.5 | Latest models | 4 | 20 | 0.20 | 5 |
| 8= | Claude Opus 5 | Legacy models | 5 | 25 | 0.50 | 6.25 |
| 8= | Claude Opus 4.8 | Legacy models | 5 | 25 | 0.50 | 6.25 |
| 8= | Claude Opus 4.7 | Legacy models | 5 | 25 | 0.50 | 6.25 |
| 8= | Claude Opus 4.6 | Legacy models | 5 | 25 | 0.50 | 6.25 |
| 8= | Claude Opus 4.5 | Legacy models | 5 | 25 | 0.50 | 6.25 |
| 13= | Claude Fable 5.1 | Latest models | 10 | 50 | 0.25 | 12.50 |
| 13= | Claude Mythos 5.1 | No card | 10 | 50 | 0.25 | 12.50 |
| 13= | Claude Fable 5 | Legacy models | 10 | 50 | 1 | 12.50 |
| 13= | Claude Mythos 5 | No card | 10 | 50 | 1 | 12.50 |
| 17= | Claude Opus 4.1 | No card | 15 | 75 | 1.50 | 18.75 |
| 17= | Claude Opus 4 | No card | 15 | 75 | 1.50 | 18.75 |
Prices: Anthropic pricing documentation, Model pricing table. Card placement: Anthropic’s pricing page. Both checked 28 September 2026.
Cache hits break the input order in two places. Anthropic sets a cache hit at 0.1x base input on most models, 0.05x on Claude Opus 5.5 and 0.025x on Claude Fable 5.1 and Claude Mythos 5.1.1
Opus 5.5 therefore reads from cache at $0.20, the same price as Sonnet 5. On cached tokens the gap between Haiku 4.5 and Opus 5.5 narrows from four times to two. Coverage of model serving economics sits on the model infrastructure desk.
What does Claude Haiku 4.5 cost per million tokens?
Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens at standard rates. A cache hit costs $0.10. A 5-minute cache write costs $1.25 and a 1-hour cache write $2.1
| Haiku 4.5 rate | $ per MTok | Anthropic basis |
|---|---|---|
| Standard input | 1 | Model pricing table |
| Standard output | 5 | Model pricing table |
| Cache hit | 0.10 | Model pricing table |
| 5-minute cache write | 1.25 | Model pricing table |
| 1-hour cache write | 2 | Model pricing table |
| Batch input | 0.50 | Batch processing table |
| Batch output | 2.50 | Batch processing table |
| Batched cache hit | 0.05 | 0.10 with the 50% batch discount applied |
Anthropic’s pricing page calls Haiku 4.5 its “Fastest, most cost-efficient model” and shows the same standard rates. It notes that its cache prices reflect a 5-minute TTL.2
Output costs five times input. A classification call with a long prompt and a one-word answer pays almost entirely for input, so caching moves its bill most. A drafting call pays mostly for output, where only batch processing or a cheaper model cuts the rate.
Which Haiku versions does Anthropic still price?
Anthropic prices two Haiku models: Claude Haiku 4.5 and Claude Haiku 3.5. Haiku 3.5 sits at the foot of the documentation table at $0.80 input and $4 output, with cache hits at $0.08 and a 1-hour cache write at $1.60.1
Anthropic’s main pricing page lists Haiku 4.5 under “Latest models”.2 Its “Legacy models” section holds Opus, Sonnet and Fable cards and no Haiku card.2
CloudPrice, a third-party price tracker, lists a Haiku 3.5 deprecation date of 4 July 2026.3 The same tracker lists Claude Haiku 3 at $0.250 input and $1.25 output, marked “Deprecated”.3
Anthropic’s documentation table has no Claude 3 Haiku row. On the evidence captured, Haiku 4.5 and Haiku 3.5 are the only Haiku versions Anthropic still prices.
The decision changes when a workload already runs on Haiku 3.5. Its rates sit 20% below Haiku 4.5 on both input and output. A price row shows a rate. It gives no date for when the model stops answering requests. Confirm Haiku 3.5 availability in the Claude Console before you budget on it.
How do prompt caching and batch processing change the Haiku bill?
A cache hit costs 0.1 times the base input price on Haiku, and a 5-minute cache write costs 1.25 times. Anthropic states that these multipliers stack with the Batch API discount.1
The Batch API processes requests asynchronously with a 50% discount on input and output tokens.1
Take a request with a 10,000-token shared prefix, such as a long system prompt, plus 500 fresh input tokens and 500 output tokens.
| Request type | Haiku 3.5 ($) | Haiku 4.5 ($) | Sonnet 5 ($) |
|---|---|---|---|
| No cache | 0.0104 | 0.0130 | 0.0260 |
| First call, 5-minute cache write | 0.0124 | 0.0155 | 0.0310 |
| First call, 1-hour cache write | 0.0184 | 0.0230 | 0.0460 |
| Later call, cache hit | 0.0032 | 0.0040 | 0.0080 |
Calculated from Anthropic’s Model pricing table, checked 28 September 2026.
On Haiku 4.5 the 5-minute write adds $0.0025 to the first call. Each later cache hit saves $0.0090 against an uncached call. The 1-hour write adds $0.0100, so it needs two hits before it pays.
Anthropic’s documentation reaches the same break-even. Caching pays off after one cache read for the 5-minute duration and after two cache reads for the 1-hour duration.1
The two discounts combine. A cached batch request on Haiku 4.5 brings input down to $0.05 per million tokens, according to Fast.io’s pricing guide.4
The constraint on batch is time. Anthropic’s cost guidance reserves the Batch API for non-time-sensitive tasks, so it fits poorly with a live chat reply.1
How much does one request cost on Haiku versus Sonnet?
A request with 2,000 input tokens and 500 output tokens costs $0.0045 on Haiku 4.5 and $0.0090 on Sonnet 5 at standard rates. Sonnet 4.6 charges $0.0135 for the same token counts, three times Haiku 4.5.
| Model | Input ($/MTok) | Output ($/MTok) | Per request, standard ($) | Per request, batch ($) | Per 1,000 requests, standard ($) |
|---|---|---|---|---|---|
| Claude Haiku 3.5 | 0.80 | 4 | 0.0036 | 0.0018 | 3.60 |
| Claude Haiku 4.5 | 1 | 5 | 0.0045 | 0.00225 | 4.50 |
| Claude Sonnet 5 | 2 | 10 | 0.0090 | 0.0045 | 9.00 |
| Claude Sonnet 4.6 | 3 | 15 | 0.0135 | 0.00675 | 13.50 |
Calculated from Anthropic’s Model pricing and Batch processing tables, checked 28 September 2026. Request: 2,000 input tokens and 500 output tokens.
The formula is plain. Multiply input tokens by the input rate and output tokens by the output rate, add the two, then divide by 1,000,000. For Haiku 4.5: (2,000 × $1 + 500 × $5) ÷ 1,000,000 = $0.0045.
Token counts differ between models for the same text. Anthropic’s documentation says Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text. Claude Sonnet 4.6 and earlier models use the previous tokenizer.1
By version number, Sonnet 5 falls under the newer tokenizer and Haiku 4.5 under the previous one. If the same text becomes 2,600 input tokens and 650 output tokens on Sonnet 5, the request costs about $0.0117. Sonnet 5’s premium over Haiku 4.5 then rises from 2 times to about 2.6 times. The exact increase depends on the content.
Tool use adds a hidden system prompt to each request. Anthropic lists 496 tokens for Haiku 4.5 and 354 tokens for Sonnet 5 when tool choice is auto or none.1
At list rates that overhead costs $0.000496 per request on Haiku 4.5 and $0.000708 on Sonnet 5. Haiku keeps its lead even with the longer tool prompt.
When is Haiku the cheapest Claude option?
Glamdring Research’s reading of Anthropic’s rates is that Haiku is the cheapest Claude line on every token price Anthropic publishes. Haiku 4.5 is the cheapest model with a card on the main pricing page. Haiku 3.5 undercuts it in the documentation table.
Per task, Haiku 4.5 loses only when a larger model finishes the job on far fewer billed tokens. Sonnet 5 ties Haiku 4.5 when it bills half as many tokens at the same input and output mix. Sonnet 4.6 ties at one third.
Retries count as tokens. If Haiku 4.5 needs a full second attempt on every request, its cost doubles and only then matches one Sonnet 5 attempt. Against Sonnet 4.6, the tie arrives at three Haiku attempts per request.
Anthropic’s own cost guidance points the same way: “Choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning”.1
Use one real workflow to test the claim. Run the same inputs through Haiku 4.5 and Sonnet 5. Count billed tokens and attempts per accepted result, then multiply by each model’s list rates. The cost drivers beyond the price list are set out in what controls AI inference cost.
Why do other published Haiku prices disagree?
Published Haiku prices differ by channel, model version and cache duration. The top Google result for “claude haiku api pricing” on 28 September 2026 quoted Haiku 4.5 at $1.10 input and $5.50 output. That figure is a provider rate for AWS Bedrock, with a cache write of $1.38 and a cache read of $0.11.5
Anthropic states that Bedrock and Google Cloud set independent regional pricing. A buyer on a cloud marketplace pays that marketplace’s rate.1
Version labels drift on trackers. CloudPrice’s version table marks Claude Haiku 4.5 “Deprecating” and Claude Haiku 3.5 “Current”.3 Anthropic’s pricing page lists Haiku 4.5 under “Latest models” on the same date.2
Tokenizer figures vary too. Fast.io says Opus 4.7 and later models produce up to 35% more tokens for the same input text.4 Anthropic’s documentation gives approximately 30%.1
What the ranking cannot tell you
The ranking measures list price per token on Anthropic’s direct API. It says nothing about output quality, speed or the tokens each model spends on a task. Anthropic’s description of Haiku 4.5 as its fastest model is a company claim.
Per-token rank and per-text cost can diverge because newer models count the same text as more tokens. Compare models on the same workload before you rely on the per-token order.
Negotiated prices sit outside the ranking. Anthropic says volume discounts may be available for high-volume users and are negotiated case by case.1
Every figure carries the checked date of 28 September 2026. Other dated cost and market work sits in the Glamdring research archive.
Frequently asked questions
How much does it cost to use the Claude API?
Anthropic bills the Claude API per token on actual monthly usage. New users receive a small amount of free credits to test the API.1
Standard rates on 28 September 2026 run from $0.80 input and $4 output per million tokens on Haiku 3.5 to $15 input and $75 output on Opus 4.1 and Opus 4. Haiku 4.5 costs $1 input and $5 output.1
Does US-only inference change Haiku’s price?
Anthropic’s documentation applies a 1.1x multiplier to US-only inference on Claude 4.6 and later models. Earlier models don’t support the inference_geo parameter and always use standard pricing. A request that includes the parameter on those models returns a 400 error.1
Haiku 4.5 and Haiku 3.5 both carry version numbers below 4.6. Anthropic’s main pricing page prints its 1.1x note under the Haiku 4.5 card without naming the models it covers.2 Confirm the behaviour in the Claude Console before you plan US-only Haiku traffic.
Can I still use Claude 3 Haiku through Anthropic?
Anthropic’s documentation table has no Claude 3 Haiku row on 28 September 2026. Its cheapest Haiku row is Claude Haiku 3.5.1
CloudPrice lists Claude Haiku 3 at $0.250 input and $1.25 output and marks it “Deprecated”.3 Treat that as a tracker record and confirm availability in the Claude Console.
Sources checked
Company-owned pages establish what a company says. They do not prove a market conclusion. Each source is dated so readers can judge each claim.
