API model string
claude-haiku-4-5Context window
200K tokens
Max output not published (standard Anthropic sync limits apply)
- Knowledge cutoff
- not published
- Released
- not published
- Modalities
- text + image in, text out
Last verified July 10, 2026
Rate limits by tier
Anthropic replaced numeric Tier 1–4 with Start, Build, Scale (+ Custom) on June 26, 2026. Monthly spend caps: Start $500, Build $1,000, Scale $200,000. As of June 2026, Haiku 4.5 carries the same RPM limits as Opus 4.x and Sonnet 4.x/5 per tier, but has a significantly higher ITPM ceiling at Scale (~4M vs ~2M for Opus/Sonnet). Haiku 4.x models share one org-level pool separate from Opus 4.x and Sonnet 4.x/5 pools. Key distinction: unlike legacy Haiku 3.5 (which carried a † exception where cache reads DID count toward ITPM), cache reads for Haiku 4.5 do NOT count toward ITPM. Figures are per Anthropic docs via third-party trackers — verify in your console.
| Tier | Requirements | RPM | TPM | RPD | Concurrent | Notes |
|---|---|---|---|---|---|---|
| Start | First API key or cumulative deposit under $40 | ~50 (per third-party trackers — confirm in your console/dashboard) | ITPM not published separately at Start tier for Haiku class; Haiku 4.5 has the highest ITPM ceiling in Anthropic's lineup at Scale (~4M); Start-tier ITPM not published | not published | not published | Entry tier. Adequate for testing and prototyping. Haiku class has highest ITPM ceiling at Scale but Start-tier ceiling is not published. Spend cap $500/month. |
| Build | Cumulative spend above ~$40; automatic promotion | not published | not published | not published | not published | Automatic on crossing threshold. Spend cap $1,000/month. |
| Scale | Cumulative deposit threshold (~$200 equivalent); spend cap $200,000/month | ~4,000 (per third-party trackers — confirm in your console/dashboard) | ~4,000,000 ITPM (Haiku class — highest in Anthropic lineup, vs ~2M for Opus/Sonnet; per third-party trackers — confirm in your console/dashboard); ~400,000–800,000 OTPM | not published | not published | Haiku 4.5 has the highest published ITPM ceiling among Anthropic models, designed for high-throughput routing, classification, and filtering. Cache reads excluded from ITPM (unlike legacy Haiku 3.5). Higher limits via Custom. |
| Custom | Contact Anthropic sales | not published | not published | not published | not published | No spend cap. Negotiated ceilings above Scale. Dedicated support SLA. |
Swipe the table sideways to see every limit column.
- 1.Token-bucket replenishment is continuous. Acceleration limits can trigger 429 on sudden spikes even within tier ceilings — especially relevant for high-frequency Haiku use cases like routing. Ramp gradually.
- 2.Cache reads for Haiku 4.5 do NOT count toward ITPM. This differs from legacy Haiku 3.5 (†) where cache reads DID count toward ITPM. Verify you are calling claude-haiku-4-5 and not a Haiku 3.5 string.
- 3.Haiku 4.x org pool is separate from Opus 4.x and Sonnet 4.x/5 pools. The ~4M ITPM ceiling at Scale is the Haiku class ceiling and does not compete with your Sonnet or Opus pool.
Limits verified against the Anthropic docs, July 10, 2026.
Token pricing
What you pay per million tokens (USD). Input and output are billed separately.
Input
$1.00
per 1M tokens
Output
$5.00
per 1M tokens
Cached input
$0.10
per 1M cached
Batch discount
−50%
on batch jobs
- Cache read = $0.10/MTok (0.1× base input $1.00) — cheapest cached-input rate among Anthropic flagship models. Cache reads excluded from ITPM for Haiku 4.5.
- 5-minute cache write = $1.25/MTok (1.25× base). 1-hour cache write = $2.00/MTok (2× base).
- Batch API (50% off): $0.50 input / $2.50 output per MTok. Cheapest Anthropic batch rate. Async within 24 hours. Max 300K output tokens per request with output-300k-2026-03-24 beta header.
- Extended thinking tokens billed as output. Disable for Haiku 4.5 — thinking defeats the low-cost fast-inference value proposition.
- US-only data residency adds 1.1× surcharge on applicable Haiku versions. Omit unless legally required.
- Web search tool adds $10 per 1,000 searches on top of token costs. Use sparingly on Haiku; cache search results in prompt context.
- No long-context surcharge on Haiku 4.5 (200K context flat). No free API tier. No first-party embeddings or fine-tuning.
Side-project chatbot
$10
per month
Assumptions
5M input tokens / 1M output tokens per month, no caching
5M × $1.00/MTok = $5 input + 1M × $5.00/MTok = $5 output = $10/mo
Mid-scale app with caching
$82.20
per month
Assumptions
60M input / 12M output per month, 70% of input served from cache
18M × $1.00/MTok = $18 (non-cached) + 42M × $0.10/MTok = $4.20 (cache reads) + 12M × $5.00/MTok = $60 output = $82.20/mo
High-volume classification pipeline
$800
per month
Assumptions
400M input / 80M output per month, no caching
400M × $1.00/MTok = $400 input + 80M × $5.00/MTok = $400 output = $800/mo
Run your own numbers
Drag your real monthly token volumes and watch the bill update live — priced against rival models at the same usage.
Estimated claude-haiku-4-5 spend
$55/mo
Input: $30
Output: $25
30M in × $1.00 + 5M out × $5.00 = $55
Same volume, priced across models
- Gemini 2.5 FlashCheapest$22
- claude-haiku-4-5This model$55
- Claude 4 Sonnet$110
- Claude 4 Opus$275
Rivals priced at their published input/output rates for the same monthly volumes. Prompt caching is model-specific, so it is applied to claude-haiku-4-5 only. Estimates for comparison; real bills vary with request shape and long-context surcharges.
claude-haiku-4-5 vs the alternatives
Claude Haiku 4.5 competes with Claude Sonnet 5 (higher capability Anthropic option), Claude Opus 4.8 (premium Anthropic option), and Gemini 2.5 Flash (Google's fastest low-cost model) on price, throughput, and latency.
| Aspect | claude-haiku-4-5 | Claude 4 Sonnet | Claude 4 Opus | Gemini 2.5 Flash |
|---|---|---|---|---|
| Context window | 200K tokens | 1M tokens | 1M tokens | 1M tokens |
| Input $/MTok | $1.00 | $2.00 (intro) | $5.00 | $0.30 |
| Output $/MTok | $5.00 | $10.00 (intro) | $25.00 | $2.50 |
| Cached input $/MTok | $0.10 | $0.20 | $0.50 | ~$0.03 |
| Batch discount | 50% | 50% | 50% | 50% |
| Max ITPM ceiling (Scale) | ~4M (highest Anthropic class) | ~2M | ~2M | not published |
| Cache reads count toward ITPM | NO (Haiku 4.5) | NO | NO | yes (context cache has storage costs) |
| Long-context surcharge | none (200K flat) | none (1M flat) | none (1M flat) | none (Flash flat) |
| Provider ecosystem / compliance | Anthropic safety/trust | Anthropic safety/trust | Anthropic safety/trust |
Swipe the table sideways to see every model.
Hitting a 429? The playbook
The exact errors you'll see
rate_limit_errortype: "rate_limit_error"HTTP 429 Too Many RequestsWhy it happens & how to fix it
RPM burst: Haiku is often used for high-frequency classification and routing, making RPM spikes common
Spread requests evenly. Token-bucket + acceleration limits fire on sudden spikes even within tier ceilings. For high-frequency routing use cases, implement a client-side token-bucket at ~80% of your tier RPM ceiling to prevent hitting server-side acceleration limits. Add sleep or queue between bursts.
ITPM saturation despite Haiku's high ~4M ceiling
Activate prompt caching. Cache reads are excluded from ITPM for Haiku 4.5 — use cache_control breakpoints on any repeated context to free ITPM headroom. This is especially effective for classification workflows with shared system prompts.
Haiku org pool shared across multiple teams draining faster than expected
Check the Rate Limits API (GET /v1/rate-limits, launched April 25, 2026) to see remaining headroom. Partition API keys across teams and monitor anthropic-ratelimit-requests-remaining and anthropic-ratelimit-tokens-remaining on each response.
OTPM exceeded on bulk Haiku generation
Route bulk generation to Batch API ($0.50/$2.50 per MTok — cheapest Anthropic batch rate). Async within 24h. This removes OTPM pressure from the real-time path and costs 50% less.
Confusion between Haiku 3.5 and Haiku 4.5 caching rules causing unexpected ITPM hits
Legacy Haiku 3.5 (†) cache reads DID count toward ITPM. Haiku 4.5 does NOT carry this exception. Verify your model string is claude-haiku-4-5, not a Haiku 3.5 string. If you migrated from 3.5, your ITPM headroom should increase significantly on 4.5.
Retry strategy
On receiving a 429, read the retry-after header (seconds to wait). Parse anthropic-ratelimit-requests-remaining versus anthropic-ratelimit-tokens-remaining to identify whether you are RPM-bound or TPM-bound. For high-frequency Haiku use cases (routing, classification), implement a client-side token-bucket at ~80% of your tier ceiling to prevent hitting server-side limits in the first place. Exponential backoff with jitter: start at 1 second, double each retry, cap at 64 seconds.
1const ANTHROPIC_API_URL = "https://api.anthropic.com/v1/messages";23// For high-frequency Haiku use cases: client-side rate limiter4let requestsThisMinute = 0;5const SAFE_RPM_CEILING = 40; // ~80% of ~50 Start RPM ceiling67async function callHaikuWithRetry(8 payload: object,9 maxRetries = 510): Promise<Response> {11 // Client-side rate limit check12 if (requestsThisMinute >= SAFE_RPM_CEILING) {13 await new Promise((r) => setTimeout(r, 1000));14 requestsThisMinute = 0;15 }16 requestsThisMinute++;1718 let attempt = 0;19 let delayMs = 1000;2021 while (attempt <= maxRetries) {22 const res = await fetch(ANTHROPIC_API_URL, {23 method: "POST",24 headers: {25 "x-api-key": process.env.ANTHROPIC_API_KEY!,26 "anthropic-version": "2023-06-01",27 "content-type": "application/json",28 },29 body: JSON.stringify({ ...payload, model: "claude-haiku-4-5" }),30 });3132 if (res.status !== 429) return res;3334 const retryAfter = res.headers.get("retry-after");35 const reqsRemaining = res.headers.get("anthropic-ratelimit-requests-remaining");36 const toksRemaining = res.headers.get("anthropic-ratelimit-tokens-remaining");3738 console.warn(39 `429 on attempt ${attempt + 1}. ` +40 `Requests left: ${reqsRemaining}, Tokens left: ${toksRemaining}. ` +41 `Retry-After: ${retryAfter}s`42 );4344 const waitMs = retryAfter45 ? parseInt(retryAfter, 10) * 100046 : delayMs + Math.random() * 500;4748 await new Promise((r) => setTimeout(r, waitMs));49 delayMs = Math.min(delayMs * 2, 64000);50 attempt++;51 }5253 throw new Error(`Exceeded max retries (${maxRetries}) on 429`);54}How to raise your limits
The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.
Start
ImmediateAutomatic on creating your first API key or depositing under $40 cumulative. No application required.
Unlocks: ~50 RPM. Spend cap $500/month. ITPM ceiling not published at Start for Haiku class, but designed for testing and low-volume prototyping.
Build
Automatic within hours to days of crossing thresholdAutomatic when cumulative spend crosses ~$40. Monitor at platform.claude.com/settings/limits.
Unlocks: RPM and ITPM scale above Start. Spend cap $1,000/month. Suitable for early production routing and classification workloads.
Scale
Automatic within days of crossing thresholdAutomatic when cumulative deposit crosses ~$200. Check platform.claude.com/settings/limits for position. Query GET /v1/rate-limits for programmatic ceiling monitoring.
Unlocks: ~4,000 RPM / ~4,000,000 ITPM (Haiku class — highest Anthropic tier) / ~400,000–800,000 OTPM (per third-party trackers — confirm in console). Spend cap $200,000/month. Haiku 4.5 designed for high-throughput routing, classification, and filtering at this level.
Custom
Days to weeks to negotiateContact Anthropic sales. Discuss volume commitments, enterprise agreements, and BAA if required.
Unlocks: No spend cap. Negotiated RPM/ITPM/OTPM above Scale. Dedicated support SLA. Enables highest-throughput bulk classification and routing deployments.
Cut your token spend
Prompt caching
−90% cost on repeated context; cache reads excluded from ITPM — dual cost and throughput benefitCache reads = $0.10/MTok (cheapest among Anthropic models) AND excluded from ITPM for Haiku 4.5. Use cache_control breakpoints on shared system prompts in classification workflows. Use 1-hr cache writes ($2.00/MTok) for long-running sessions. This is especially powerful at scale for Haiku's primary use case of repeated-context classification.
Batch API for bulk classification
50% off = $0.50/$2.50 per MTok — cheapest Anthropic batch rateRoute bulk tagging, extraction, and classification jobs to Batch API. Async within 24h. Supports up to 300K output tokens per request with output-300k-2026-03-24 beta header. Removes OTPM pressure from real-time paths.
Client-side token-bucket
Eliminates acceleration-triggered 429s before they hit server-side limitsFor high-frequency routing use cases, implement a client-side request queue at ~80% of your tier's RPM ceiling. This smooths traffic curves and prevents server-side acceleration limits from firing on sudden bursts.
Model routing: Haiku for triage, escalate to Sonnet/Opus
Primary cost optimization pattern for multi-tier Anthropic setups; Haiku is 5× cheaper input than OpusUse Haiku 4.5 for fast classification, routing, summarization, and filtering tasks. Escalate to Sonnet 5 ($2/$10) or Opus 4.8 ($5/$25) only when task complexity warrants deeper reasoning. Haiku at $1/$5 vs Sonnet at $2/$10 saves 50% on applicable tasks.
Ramp gradually
Avoids acceleration-triggered 429s even at Haiku's high ITPM ceilingAcceleration limits apply to Haiku despite its high ITPM ceiling. Smooth request curves prevent acceleration-triggered 429s, especially important for high-frequency classification use cases that tend to ramp from 0 to full speed instantly.
Disable extended thinking on Haiku
Preserves Haiku's value proposition of low-cost fast inferenceExtended thinking tokens bill as output and multiply costs 3–5×. Haiku 4.5's value is low-cost, high-throughput inference. Enable thinking only on Sonnet or Opus, not on Haiku routing/classification layers.
Route longer contexts to Sonnet 5
Haiku's 200K context is sufficient for most routing tasks; escalation prevents context errorsHaiku 4.5 has 200K context vs 1M for Sonnet 5 and Opus 4.8. For documents or conversations exceeding 200K tokens, route to Sonnet 5. This avoids context-window errors and keeps Haiku on tasks it handles optimally.
Use web search tool sparingly on Haiku
Avoids $10/1K searches surcharge dominating Haiku's low per-token costThe web search tool charges $10 per 1,000 searches on top of token costs. For Haiku at $1/$5 per MTok, a single search costs more than processing 10,000 tokens. Cache search results in prompt context and avoid per-request web searches on Haiku.
Frequently asked questions
Is Claude 3 Haiku deprecated?
Yes. Claude 3 Haiku was officially deprecated on March 2, 2026. New projects should not use claude-3-haiku-* API strings. The recommended replacement is Claude Haiku 4.5 (claude-haiku-4-5) — cheaper, with 200K context and no ITPM exception on cache reads. As of July 10, 2026, no hard EOL date has been announced, but the model is not recommended.
What is the difference between Claude 3 Haiku and Haiku 4.5?
Claude 3 Haiku is deprecated (March 2, 2026). Haiku 4.5 (claude-haiku-4-5) is the current GA model, cheaper at $1.00/$5.00 per MTok, with 200K context. Critically, unlike Haiku 3.5 (a separate intermediate model that carried a † exception), Haiku 4.5 cache reads do NOT count toward ITPM — a significant throughput improvement for caching-heavy workflows.
How do I increase Claude Haiku 4.5 rate limits?
Rate limits increase automatically: ~$40 cumulative spend promotes to Build; ~$200 cumulative unlocks Scale (~4,000 RPM / ~4M ITPM — the highest Anthropic ITPM ceiling, per third-party trackers). Above Scale, contact Anthropic sales for Custom tier. Monitor at platform.claude.com/settings/limits or query GET /v1/rate-limits programmatically.
Is the Claude Haiku 4.5 API free?
No. There is no free tier. All token usage is billed from the Start tier. Standard pricing is $1.00 input / $5.00 output per MTok — Anthropic's lowest published rate. Batch API offers 50% off at $0.50/$2.50 per MTok.
How does Claude Haiku 4.5 compare to Gemini 2.5 Flash pricing?
Gemini 2.5 Flash is cheaper on raw token cost at approximately $0.30 input / $2.50 output per MTok, versus Haiku 4.5 at $1.00/$5.00. Haiku 4.5 offers Anthropic's safety/trust ecosystem and a very high ITPM ceiling (~4M at Scale). Gemini 2.5 Flash has a 1M context window versus Haiku's 200K. Choose based on provider compliance requirements and whether your workload fits within Haiku's 200K context.
Do cache reads count toward Claude Haiku 4.5 ITPM limits?
No — cache reads do NOT count toward ITPM for Haiku 4.5. This is an important distinction from the legacy Haiku 3.5 model, which carried a † exception where cache reads DID count toward ITPM. If you are migrating from Haiku 3.5, your effective ITPM throughput should increase significantly on Haiku 4.5 with identical caching patterns.
What does Claude API rate_limit_error mean for Haiku?
rate_limit_error (HTTP 429) on Haiku means you hit RPM, ITPM, or OTPM limits. High-frequency routing and classification use cases are most likely to hit RPM acceleration limits. Check anthropic-ratelimit-requests-remaining vs anthropic-ratelimit-tokens-remaining to identify which dimension hit. Read retry-after for seconds to wait. For persistent RPM issues, implement a client-side token-bucket at ~80% of your tier RPM ceiling.
How can I reduce Claude API costs using model routing?
The primary pattern for multi-tier Anthropic setups: use Haiku 4.5 ($1/$5 per MTok) for routing, classification, summarization, and filtering. Escalate to Sonnet 5 ($2/$10 intro) for generation tasks requiring moderate reasoning, and Opus 4.8 ($5/$25) only for tasks requiring maximum reasoning depth. Haiku is 5× cheaper than Opus on input. RapidDev helps teams design and implement model routing architectures to reduce Anthropic API spend — reach out at rapidevelopers.com/contact for a free scoping call.
What is the Haiku 4.5 context window and when does it matter?
Claude Haiku 4.5 has a 200K token context window, compared to 1M for Sonnet 5 and Opus 4.8. For most classification, routing, and summarization tasks, 200K is more than sufficient. When processing very long documents, full conversation histories, or large codebases that exceed 200K tokens, route to Sonnet 5 or Opus 4.8. Haiku 4.5 does not carry any long-context surcharge — the 200K window is available at the flat $1/$5 rate.
We build AI apps that don't hit rate limits
- Retry, backoff & caching built in
- Multi-provider fallback routing
- Fixed price, you own the code
30-min call. No commitment.