PRICING

Claude API Pricing: List Rates vs Live Quotes for 13 Models

GPU Router teamSeptember 30, 202610 min read

Claude API pricing comes down to two choices: which Claude model you call, and where you buy the tokens. For models still available on Anthropic's API, list prices run from $1 per million input tokens for Claude Haiku 4.5 up to $10 for Claude Fable 5.1. Every one of those models charges five times as much for output as for input.

GPU Router is our answer to the second choice. It's an OpenAI-compatible API that sends each request to the lowest-cost eligible route for the exact Claude model you name. On the day we checked, all 13 Claude models we route (the -fast variants aside) had standard quotes below Anthropic's list price, from 17.1% to 72.7% lower. If your code already calls Chat Completions through the OpenAI SDK, switching means changing the base URL, the API key and the model ID.

On September 30 at 18:20 UTC we pulled our live catalog and set every Claude quote beside Anthropic's list price. That snapshot is below, along with the receipt from a real request we sent through it. Treat the numbers as a snapshot: while we were writing, the Claude Opus 5.5 input quote moved from $1.0920 to $1.680 per million tokens in about 11 minutes.

How much does the Claude API cost per million tokens?

At Anthropic's list prices, Claude costs $1 input and $5 output per million tokens for Haiku 4.5, $2 and $10 for Sonnet 5 and Sonnet 5.5, $4 and $20 for Opus 5.5, and $10 and $50 for Fable 5.1. In our snapshot, GPU Router quoted the same models between 17.1% and 72.7% below those rates.

Live prices retrieved from gpurouter.io/catalog/models on September 30, 2026 at 18:20 UTC. USD per 1M tokens. Prices change.

Older Claude models are still worth pricing, because some of them are quoted lower than the newest release in the same tier:

List prices come from Anthropic's pricing page (checked September 30, 2026), and they match the direct list prices in our catalog. Our quotes include our routing fee; savings are capped at 95%. They're estimates, not a bill: the final charge depends on the route that serves the request and the tokens you actually use. "Below list" is (1 − our price ÷ list price) × 100 on 1M input plus 1M output tokens.

We left out the four -fast model IDs in our catalog. They're separate listings from the models above and deserve their own comparison.

Which Claude model is cheapest per token?

At list price, Claude Haiku 4.5 is the cheapest Claude model still available on Anthropic's API, at $1 input and $5 output per million tokens. Through GPU Router that day, Claude Sonnet 5 was cheaper per token, at $0.5460 and $2.730, below Haiku's routed quote of $0.735 and $3.68. In that snapshot, the higher-tier model was also the cheaper call.

Discounts come from unused capacity: when teams reserve or prepay for more than they end up using, suppliers can offer the difference for less. How much exists depends on each supplier and changes over time. Sonnet 5's cheapest eligible route sat 72.7% below list, while Haiku 4.5's best route was 26.4% below.

Per token isn't the whole story, though. Anthropic says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, and Haiku 4.5 predates it. With 30% more tokens, Sonnet 5's 1M-input-plus-1M-output cost rises from $3.28 to about $4.26, against $4.41 for Haiku. Sonnet 5 still came out ahead, but only just.

Sonnet 5 is also cheaper than it was meant to be. Anthropic launched it at $2 / $10 as an introductory price and has since made that the standard rate, canceling the planned September 1 increase to $3 / $15. You can put the two models side by side on our comparison page. And if a model router picks the Claude tier for each prompt, give it the prices you actually pay, not list prices.

What does one Claude request actually cost?

We sent one short prompt to Claude Sonnet 5 through GPU Router at 18:22 UTC that day. It used 27 input tokens and 64 output tokens, and the receipt charged $0.000177. At Anthropic's list price those tokens cost $0.000694, so this request came in 74.5% below list with our routing fee included.

The receipt records what the request was actually billed, a little under what our 18:20 quote implied for the same tokens ($0.000190). Activity in your account records each request's timings, and GET /v1/receipts/{reservationId} returns its receipt from the API. Our latency docs cover both.

A single request is tiny, so let's scale it up. Say a coding assistant sends 30M input tokens in a month and gets 3M output tokens back. At base rates, before any caching:

The table holds token counts equal across models; with the tokenizer difference above, the same work on Sonnet 5 would come to about $31.94 at our quote, still under Haiku's $33.08.

Do prompt caching and the Batch API still matter?

Yes. At Anthropic's rates, cache hits cost 10% of the base input price (5% on Opus 5.5, 2.5% on Fable 5.1), and the Batch API halves both input and output prices for asynchronous work. Our quotes above are base token rates, so a workload that's mostly cache hits at Anthropic can close a good part of the gap.

Here's the same month with caching. If 80% of those 30M input tokens were cache hits on Sonnet 5, Anthropic's bill would drop from $90.00 to about $46.80, ignoring the cost of writing the cache. That's still above our $24.57 base-rate quote, but the gap is much smaller. Writing the cache isn't free either: Anthropic charges 1.25x the base input price for a five-minute cache write and 2x for a one-hour write.

If your prompts reuse large prefixes, run that arithmetic with your own cache-hit rate before you decide. Anthropic's prompt caching guide and batch processing docs have the full multipliers.

How to call Claude through GPU Router

GPU Router uses the OpenAI Chat Completions format, so you call Claude with the OpenAI SDK. Point base_url at https://gpurouter.io/v1, pass your GPU Router key, and use a model ID with the anthropic/ prefix. The quickstart walks through the same setup, and our AI SDK OpenAI-compatible guide covers it for TypeScript apps. We ran this script on September 30 at 18:21 UTC:

import os

from openai import OpenAI

client = OpenAI(
    base_url="https://gpurouter.io/v1",
    api_key=os.environ["GPU_ROUTER_API_KEY"],
)

response = client.chat.completions.create(
    model="anthropic/claude-sonnet-5",
    messages=[
        {"role": "user", "content": "In one sentence: why do output tokens cost more than input tokens?"}
    ],
    max_tokens=64,
)

print(response.choices[0].message.content)
print("input tokens:", response.usage.prompt_tokens, "output tokens:", response.usage.completion_tokens)

It printed the start of an answer, cut off mid-sentence by the 64-token cap we set, then input tokens: 27 output tokens: 64. Raise max_tokens when you want complete replies. The curl version below is the request behind the receipt above:

curl https://gpurouter.io/v1/chat/completions \
  -H "Authorization: Bearer $GPU_ROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-5",
    "messages": [{"role": "user", "content": "In one sentence: why do output tokens cost more than input tokens?"}],
    "max_tokens": 64
  }'

A few things to know before you switch:

• The Anthropic SDK won't work here. We don't offer the Anthropic Messages API, so the Anthropic SDK and Claude Code can't be pointed at our base URL. For coding tools, we publish setup guides for Cursor, Zed, Continue, Cline, OpenCode and Aider.
• Zero data retention is per request. Add extra_body={"zdr": True} to the Python call (or top-level "zdr": true in JSON) on every request. ZDR has its own quotes: Sonnet 5 was $1.575 / $7.875 with ZDR in our snapshot. If no ZDR route can serve the request, you get HTTP 503 privacy_route_unavailable. The ZDR docs and the ZDR model list cover the details.
• Keys and credit. Create a key on the API keys page; new accounts get $1 in free prepaid credits. Authentication explains how keys and Bearer auth work.
• Coming from OpenRouter? Our OpenRouter alternatives comparison prices Claude Opus 5.5 and Claude Sonnet 5 on both services under the same anthropic/ model IDs.

When GPU Router won't save you much

Not every Claude model gets a deep discount. Claude Sonnet 5.5 had a single eligible discounted route in our snapshot, quoted 17.1% below list. Its only ZDR route was quoted at $2.276 / $11.380, above Anthropic's $2 / $10.

Anthropic's own discounts can also win. Its Batch API and prompt caching discounts stack, so the Sonnet 5 month above, run as batch jobs with 80% cache hits, would cost about $23.40 direct (again ignoring cache writes). That's below our $24.57 base-rate quote. For asynchronous work that reuses long prompts, price both before you switch.

If the discounted routes for a model fail, the Reference provider at list price is the last fallback, plus our routing fee. The request can still complete, just without a discount. Failures before the response starts can retry, but a response already in progress isn't restarted. Routing describes the order in detail.

Quotes also move, so if a request isn't worth sending at full price, turn on minimum quoted savings in Settings. When no eligible quote meets your minimum after fees, paid requests stop with HTTP 503 minimum_savings_not_met instead of going out at list. It checks quotes before sending and doesn't guarantee the final bill. It's off by default.

Check today's Claude prices

Every quote above is from one afternoon. To see today's, open the model you care about in our model catalog, which lists live route prices. When you're ready to try one, create a key at /router/api-keys and spend the free $1 on the request above.