Cheap Claude API at 14.2% of list: prices and setup

Most "cheap Claude API" pages quote a discount and stop. This one shows the arithmetic: the list price, the route the request is billed on, and the resulting percentage. Every figure below was read from our live price catalogue on 22 October 2026 and cross-checked against Anthropic's published rates.

What a cheap Claude API actually is

There are three ways to get Claude through an API, and only one of them is a discount on the same endpoint:

A gateway does not host Claude. It holds upstream capacity and forwards your request to it, which is why the same model ID answers at a fraction of the price. The trade is real: you get the model, you do not get Anthropic's account relationship. Any page that hides that trade is selling you something else.

Claude API pricing: list vs what you pay

<div class="pricebox"><table><thead><tr><th>Model</th><th>Official list / 1M</th><th>Token Master / 1M</th><th>You pay</th></tr></thead><tbody><tr><td><code>claude-opus-4-8</code></td><td>$5.00 in / $25.00 out</td><td>$0.71 in / $3.55 out</td><td>14.2%</td></tr><tr><td><code>claude-sonnet-4-6</code></td><td>$3.00 in / $15.00 out</td><td>$0.43 in / $2.13 out</td><td>14.2%</td></tr><tr><td><code>claude-sonnet-5</code></td><td>$2.00 in / $10.00 out</td><td>$0.28 in / $1.42 out</td><td>14.2%</td></tr><tr><td><code>claude-haiku-4-5</code></td><td>$1.00 in / $5.00 out</td><td>$0.14 in / $0.71 out</td><td>14.2%</td></tr></tbody></table><p>Prices read from the live catalogue on 2026-10-22.</p></div>

The mechanism behind the last column: our Claude route bills each request at 0.71 × the upstream list price × a $0.20 credit, which lands at 14.2% of list. Nothing is rounded in your favour on the page and against you on the invoice — the table is generated from the same catalogue the billing reads, and the percentage is computed, not typed.

Two things the table does not show. Thinking variants (claude-opus-4-7-thinking, claude-opus-5-thinking) are billed by our upstream at 1.5× the base model's list, so they cost $1.07 / $7.46 per million instead of $0.71 / $3.55 — still 14.2% of *their* list. And list prices move: Anthropic has changed them twice in 2026, which is why every table here carries a date.

Where the 86% comes from, and what you give up

Three things make the discount possible, and none of them is charity:

• Volume. Upstream sellers buy capacity in bulk and resell it by the credit. We buy from them, add

a fixed credit price, and pass the rest through.

• Pooled capacity. Requests are spread across upstream accounts, so a single rate limit rarely

stops a single customer. Failover between upstreams is automatic.

• Prepaid. You pay before you spend. No card on file, no invoicing, no collections risk — that is

worth several points of margin to the seller.

What you do not get, stated plainly:

• No Batch API, Files API or embeddings on this endpoint. Chat, messages, streaming, tool calling

and vision only.

• No SLA and no enterprise support contract. Support is email and the dashboard.
• No Anthropic account relationship. You are a Token Master customer, not an Anthropic one.
• Upstream risk. Supply can degrade or change without notice. Credit does not expire, but the

route serving a specific model can move.

If any of those four is a hard requirement, pay list price. The point of a cheap Claude API is cost, not procurement comfort.

Connect Claude Code in three lines

Claude Code reads three environment variables and keeps working exactly as before:

export ANTHROPIC_BASE_URL=https://cheapapi.site
export ANTHROPIC_AUTH_TOKEN=sk-tm-your-key
export ANTHROPIC_MODEL=claude-sonnet-4-6
export ANTHROPIC_SMALL_FAST_MODEL=claude-haiku-4-5
claude

Do not add /v1 to an Anthropic base URL — the SDK appends /v1/messages itself and you get a 404. To make it permanent, put the same four variables under env in ~/.claude/settings.json. Full reference in the docs.

That is the whole migration. The claude code api key you were using before is replaced by one sk-tm- key that also works for every other model on the gateway, and the Claude Code session no longer bills against a $20 subscription with a usage window.

The SDKs, Cursor, Cline and Codex CLI

Both wire formats are served from the same key, which is why the same account covers Anthropic-shaped and OpenAI-shaped tools.

import anthropic

client = anthropic.Anthropic(api_key="sk-tm-your-key", base_url="https://cheapapi.site")
msg = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarise this diff in one line."}],
)
print(msg.content[0].text)

from openai import OpenAI

client = OpenAI(api_key="sk-tm-your-key", base_url="https://cheapapi.site/v1")
print(client.chat.completions.create(
    model="claude-sonnet-4-6",
    messages=[{"role": "user", "content": "Same model, OpenAI shape."}],
).choices[0].message.content)

export OPENAI_BASE_URL=https://cheapapi.site/v1
export OPENAI_API_KEY=sk-tm-your-key
codex

The OpenAI-compatible path is the one that makes this an anthropic api cheap option rather than an Anthropic-only one: point any OpenAI client at /v1 and switch the model string to a Claude ID when you want Claude. Per-tool notes live in the SDK section and the Anthropic section.

Switching an existing key without breaking anything

If an sk-ant- key is already in production, the migration is three edits and one test:

1. Replace the base URL with https://cheapapi.site (or ANTHROPIC_BASE_URL in Claude Code).
2. Replace the key with an sk-tm- key from the dashboard.
3. Keep the model string you already use — claude-sonnet-4-6 and claude-opus-5 both route here.

What breaks, in order of likelihood: calls to /v1/files, /v1/batches and /v1/embeddings (not offered here, keep those on Anthropic), code that reads organization_id from responses, and spend alerts wired to Anthropic's billing page — point those at the gateway dashboard instead.

Test before cutting over: send the same request to both endpoints and compare response.model, usage.input_tokens and usage.output_tokens. If the counts match, the only thing that changed is the invoice. Rollback is restoring the old variables; no state moves, because credit is prepaid rather than metered against an account. The team behind this page runs the gateway and reads its billing logs for a living — who we are — and every number here comes from the same catalogue the billing reads.

What 500 chats actually cost

Token counts, not adjectives. A "chat" here is 4,000 input and 1,500 output tokens — a long system prompt plus a real question.

Credit is denominated in list-price usage, not in dollars charged: $10 buys $50 of it, and the first two rows above come to $44.25 of usage — $6.28 actually charged. That is what claude api credits means here, and the same catalogue prices the packs and per-key limits.

Prompt caching on a discounted route

Caching survives the gateway, and it compounds the discount rather than replacing it. On Opus 5 a cached input token costs $0.07 per million here against $0.50 on Anthropic's list — the same 14.2% ratio applied to the cache-read rate. The rule of thumb does not change: cache anything you resend more than twice in an hour, and do not bother below roughly 1,000 tokens of stable prefix, because the write premium eats the saving. Details in Anthropic's caching documentation.

When a cheap Claude API is the wrong choice

• You need a signed contract, a DPA you negotiated, or an invoice your accountant accepts without

questions.

• You need Batch, Files, embeddings, or a specific data-residency commitment.
• Your workload is latency-critical and you cannot tolerate an upstream failover.
• You are building something a regulator will audit and "resold capacity" is a finding.

For everything else — side projects, agent loops, internal tools, a coding assistant that runs all day — the list price is the wrong price to pay.

FAQ

Is this the same Claude model?

Yes. The model ID you request is the model ID routed, and response.model echoes it. What differs is who holds the upstream account and what you pay per token, not the weights.

Do I need an Anthropic account or a card?

No. Sign up with an email, create a key, and top up credit — card and crypto rails are listed on the payments page.

Will Claude Code, Cursor and Cline work without changes?

Yes — they only need a base URL and a key. Claude Code uses the Anthropic variables above; Cursor, Cline and Codex CLI use the OpenAI-compatible /v1 endpoint.

What are the rate limits?

Per-key RPM limits depend on the pack you buy, and each key can be capped separately. See /#packages for the current ladder.

What is the cheapest Claude API for small workloads?

For anything under a few million tokens a month, a prepaid gateway is the cheapest option because there is no subscription floor: 500 chats cost $2.45 here, and $0 if you do not use them. The cheapest claude api comparison only flips in Anthropic's favour at enterprise volume with negotiated rates.

Can I get a refund?

Credit is final except where the fault is ours — the exact wording is in the refund policy. Read it before topping up; it is short.

Sources

Prices on this page were checked on 22 October 2026 against Anthropic's API pricing, and the setup steps follow Claude Code's settings reference. Token Master is a resale gateway, not an Anthropic partner, and does not claim official status. The page is maintained by the team that operates the gateway and is re-checked every time a list price changes.