Claude API Pricing 2026: Opus, Sonnet and Haiku 5.5

Anthropic's list prices for the Claude API are $4 input / $20 output per 1M tokens for Opus 5.5, $2 / $10 for Sonnet 5.5, and $0.10 / $0.50 for Haiku 5.5 on prompts up to 100,000 tokens. Haiku 5.5 jumps to $0.50 / $2.50 once a prompt goes over that threshold. On Orux AI the same three models cost 380 / 1,900, 190 / 950 and 10.8 / 54 Credits per 1M tokens ($1.90 / $9.50, $0.95 / $4.75 and $0.054 / $0.27), roughly 47.5–56% of the official rate.

Below are the full tables, including cache writes and cache reads, plus the two places where the headline numbers mislead: Haiku's 100K cliff and the Batch API. Prices as of Oct 29, 2026.

Claude API prices at a glance

Official Anthropic prices, per 1M tokens, from the Claude pricing page:

Swipe horizontally to view the full table.

Orux AI prices for the same models, per 1M tokens. 1 Credit = $0.005, so 200 Credits = $1:

Swipe horizontally to view the full table.

Two details are easy to miss. First, on Opus 5.5 and Sonnet 5.5 Anthropic prices a cache read at 0.05x the input price, not the 0.1x used by older models. That makes caching even more valuable on those two. Second, every Orux AI chat request has a minimum charge of 0.001 Credits, which only matters for very short calls.

Haiku 5.5's 100K prompt threshold

Haiku 5.5 is the only model in this family priced by prompt length. Anthropic counts every input token toward the threshold, including cache reads and cache writes, and prices each request on its own. A request one token over the line pays the higher rate for the entire request, not just the excess.

The cliff is steep. With 1,000 output tokens:

Swipe horizontally to view the full table.

One extra token makes the request about 5.2 times more expensive in both places. If you feed Haiku long documents, trim or chunk them to stay under 100,000 input tokens, and remember that a cached system prompt still counts. Orux AI applies the same boundary: prompts up to and including 100,000 tokens use the lower tier.

Opus 5.5 and Sonnet 5.5 have no long-prompt surcharge on Anthropic's price list, and Orux AI publishes a single rate for each of them as well.

Prompt caching: when it pays off

Caching works by paying a little more once to store a prompt prefix, then paying much less every time a later request reuses it. A 5-minute cache write costs 1.25x the input price and a 1-hour write costs 2x. On Opus 5.5 and Sonnet 5.5 a read costs 0.05x.

Worked example on Sonnet 5.5: a 10,000-token system prompt reused across 1,000 requests, with traffic steady enough that the 5-minute cache stays warm.

Swipe horizontally to view the full table.

The prefix cost drops by 93–95% in both columns. The arithmetic for break-even is simple. A 5-minute write costs an extra 0.25x the input price; each read saves 0.95x on Opus and Sonnet 5.5 and 0.9x on Haiku. So a 5-minute cache pays for itself on the first reuse. A 1-hour write costs an extra 1x, so it needs two reuses.

Caching does not help with output tokens, and it does not help if your prefix changes on every call. Put the stable parts first — system prompt, tool definitions, reference documents — and the per-request content last.

Batch API: 50% off direct, not available on Orux AI

Anthropic's Batch API processes requests asynchronously at half the standard price, and the discount stacks with caching. Orux AI does not offer batch processing for any model at the moment; every request is a normal synchronous or streaming call.

How the two compare, per 1M tokens:

Swipe horizontally to view the full table.

If you run large offline Haiku jobs that can wait for batch completion, the direct Batch API is cheaper than Orux AI. For Opus and Sonnet, Orux AI's standard rate is slightly below Anthropic's batch rate, and you get responses immediately.

Cost per task: three workloads

Per-million rates are hard to reason about, so here are three typical request shapes, priced per 1,000 requests without caching:

Swipe horizontally to view the full table.

In each cell, the first figure is Orux AI and the second is Anthropic's list price.

The coding row is where caching earns its keep. Agent turns resend the same repository context again and again. If 45,000 of those 50,000 input tokens are cache reads on Sonnet 5.5, the cost per 1,000 turns falls from 13,300 Credits to 5,290 Credits ($26.45), and from $140 to $54.50 at official prices.

The model choice matters more than any discount. Opus 5.5 costs exactly twice Sonnet 5.5 and roughly 35 times Haiku 5.5 on the support workload. Route easy traffic to Haiku and save Opus for the requests that actually fail on Sonnet.

Python cost calculator

Orux AI exposes Claude through the OpenAI-compatible endpoint at https://orux.top/api/v1, so the OpenAI Python SDK works with only base_url changed. The script below calls a model and converts the returned usage into Credits using the published rates, including Haiku's two tiers.

import os
from openai import OpenAI

client = OpenAI(base_url="https://orux.top/api/v1", api_key=os.environ["ORUX_API_KEY"])

# Credits per 1M tokens on Orux AI, checked 2026-10-29
# (input, output, cache_read, cache_write_5m)
RATES = {
    "claude-opus-5-5":   [(None, (380, 1900, 22, 475))],
    "claude-sonnet-5-5": [(None, (190, 950, 12, 237.5))],
    "claude-haiku-5-5":  [(100_000, (10.8, 54, 1.1, 13.75)),   # prompt <= 100K
                          (None,    (56, 280, 5.6, 70))],       # prompt > 100K
}
USD_PER_CREDIT = 0.005

def credits_for(model: str, usage) -> float:
    details = getattr(usage, "prompt_tokens_details", None)
    cached = getattr(details, "cached_tokens", 0) or 0
    written = getattr(details, "cache_creation_tokens", 0) or 0
    prompt = usage.prompt_tokens                      # includes cached + written
    for limit, (inp, out, read, write) in RATES[model]:
        if limit is None or prompt <= limit:
            break
    fresh = prompt - cached - written
    total = (fresh * inp + cached * read + written * write
             + usage.completion_tokens * out) / 1e6
    return max(round(total, 6), 0.001)                # 0.001 Credit minimum per request

model = "claude-haiku-5-5"
resp = client.chat.completions.create(
    model=model,
    messages=[{"role": "user", "content": "Classify this ticket as billing, bug or other: "
                                          "'I was charged twice for October.'"}],
    max_tokens=50,
)
c = credits_for(model, resp.usage)
print(resp.choices[0].message.content)
print(f"{resp.usage.prompt_tokens} in / {resp.usage.completion_tokens} out "
      f"= {c} Credits (${c * USD_PER_CREDIT:.6f})")

Treat the result as an estimate. The authoritative charge is the one recorded in your Orux AI usage history, and cache writes in this sketch are priced at the 5-minute rate. Requests that fail are not charged at all.

To compare the same request across vendors, swap the model string. The Claude Sonnet 5.5 API quickstart covers JSON output and validation in more depth, and the Claude Sonnet 5.5 vs GPT-6 Luna vs Gemini 4 Argon comparison runs the same prompt across three vendors.

Which Claude model for which job

Haiku 5.5 suits classification, extraction, routing and short support replies, as long as prompts stay under 100K tokens. At 48.6 Credits per 1,000 support replies it is cheap enough to run on every inbound message.

Sonnet 5.5 is the default for drafting, debugging and multi-step work. It costs 17.6 times Haiku on the same request, so it is worth measuring whether Haiku fails before upgrading.

Opus 5.5 is for long, judgement-heavy tasks where a wrong answer costs more than the tokens: large refactors, detailed analysis, hard reasoning. At twice the Sonnet price, use it selectively.

If price is the main constraint and Claude specifically is not required, our cheapest LLM API comparison ranks 14 models on the same workload.

FAQ

Q.How much does the Claude API cost per million tokens?

Anthropic lists Opus 5.5 at $4 input and $20 output, Sonnet 5.5 at $2 and $10, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100K tokens ($0.50 and $2.50 above). On Orux AI the rates are 380/1,900, 190/950 and 10.8/54 Credits, as of Oct 29, 2026.

Q.Is Claude Haiku 5.5 cheaper than GPT-6 Luna?

Not on Orux AI. GPT-6 Luna costs 5.4 Credits input and 27 Credits output per 1M tokens for prompts under 270K, exactly half of Haiku 5.5's 10.8 and 54. At official prices both list at $0.10 and $0.50. Quality differs by task, so test both on your own prompts before switching.

Q.Does prompt caching reduce Claude API cost?

Yes, when a long prefix repeats. A 5-minute cache write costs 1.25x input and a read on Opus or Sonnet 5.5 costs 0.05x, so the cache pays off on the first reuse. In our Sonnet example, a 10,000-token system prompt reused 1,000 times fell from 1,900 to about 122 Credits.

Q.Is there a free tier for the Claude API?

Anthropic's pricing FAQ says new users get a small amount of free credits for testing; there is no ongoing free tier for production use. Orux AI has no subscription or monthly fee: you top up a prepaid balance and pay per request, and failed requests are not charged.

Q.How is Orux AI billing different from Anthropic's?

Orux AI bills in Credits (1 Credit = $0.005) from a prepaid balance, at 47.5–56% of Anthropic's list prices for Claude 5.5 models. One key also covers GPT, Gemini, DeepSeek, image and video models. Orux AI does not offer the Batch API, so Anthropic's batch rate is cheaper for large offline Haiku jobs.

Sources and price check

Prices checked on 2026-10-29. Official prices, cache multipliers, the Haiku 5.5 threshold rules and Batch API rates: Anthropic Claude pricing. Orux AI rates: Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5 and GPT-6 Luna. Re-check the model pages before budgeting; prices can change.