The cheapest way to call Claude, GPT & Gemini APIs

What “cheapest” should actually mean#

A lower price per token is only a saving if the response quality, context window and tool-calling behavior match what the official API would return. Chasing the lowest number without checking integrity is how people end up on a relay that quietly serves a smaller or quantized model. So treat price and verifiability as one decision, not two.

Where a gateway lowers your real cost#

• One bill, no per-vendor overhead. No separate Anthropic, OpenAI and Google accounts, cards, and monthly reconciliation.
• Route each task to the right-priced model. Use a cheaper model for drafts and a frontier model only when it matters — all under one key.
• Add Chinese-made models without extra hurdles. DeepSeek, Kimi and Qwen are often painful to sign up for from abroad; here they're behind the same key.
• One currency, one balance. Every payment method tops up at a flat 1 RMB = $1 of credit, and every model draws from that one balance at its own USD rate — read yours on the pricing page.

What it costs#

Work out which model is actually cheapest for your job#

The headline rate per 1M tokens decides surprisingly little. What you pay is that rate multiplied by the tokens your workload really burns, and the second number is where estimates go wrong by an order of magnitude:

• Output usually costs several times input. A concise model can beat a nominally cheaper one that rambles. Compare cost per finished task, not per 1M tokens.
• Agent loops re-send the transcript. A twenty-turn Cline or Aider session bills far more than the prompt you typed, because each turn ships the accumulated context again.
• Reasoning tokens are billed. A cheap reasoning model that thinks for 3,000 tokens can lose to a pricier one that answers in 200. And each family exposes its thinking differently — GLM shares the max_tokens budget, DeepSeek returns a separate field, MiniMax needs a flag, Grok counts it in usage, MiMo 400s if the trace is dropped — see the per-family pages.
• Retries and failed tool calls are billed too. Success rate is a cost variable, which is exactly why it is worth benchmarking.
• A wrong answer costs twice. If a cheaper model needs a second pass to fix its output, you paid for two calls plus your own time.

# What a call costs = tokens actually spent x your model's live rate.
# Read the token side here; read the rate side from the pricing page.
from openai import OpenAI

client = OpenAI(base_url="https://daoxe.com/v1", api_key="YOUR_DAOXE_KEY")
r = client.chat.completions.create(
    model="YOUR_EXACT_MODEL_ID",
    temperature=0,
    max_tokens=64,
    messages=[{"role": "user", "content": "Summarize this in one line: ..."}],
)
u = r.usage
print(u.prompt_tokens, u.completion_tokens, u.total_tokens)
# Measure a REAL task, not a one-shot prompt: an agent loop re-sends the whole
# transcript every turn, so a 20-turn session bills far more than it looks.

How to call each model family#

Every family is reachable through the same base URL and key. Claude is available through both the OpenAI-compatible path and the native Anthropic Messages path — use Messages for Claude Code and Anthropic SDKs.

Quick start (60 seconds)#

1. List your models. Copy an exact ID for the family you want.
2. Send one small request. Keep max_tokens tiny to keep the test cheap.
3. Compare quality & price. Against the official API for the same prompt, at temperature=0.

export DAOXE_API_KEY="your_api_key"

curl --fail-with-body --show-error --silent \
  https://daoxe.com/v1/models \
  -H "Authorization: Bearer ${DAOXE_API_KEY}"

curl --fail-with-body --show-error --silent \
  https://daoxe.com/v1/chat/completions \
  -H "Authorization: Bearer ${DAOXE_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "YOUR_EXACT_MODEL_ID",
    "max_tokens": 64,
    "messages": [{"role": "user", "content": "Say hello in one sentence."}]
  }'

The catch nobody mentions#

Some cheap endpoints look great for a week, then quietly degrade — a smaller model, aggressive quantization, or a truncated context window. The fix isn't faith, it's measurement: run a fixed probe set on a schedule and diff against the official API. That's exactly what the open benchmark and the model-swapping guide are for.

Frequently asked questions#