Self-Hosted LiteLLM Proxy vs Requesty, Portkey and a Flat-Rate Gateway: The True Monthly Cost

"LiteLLM is free, so why pay for a gateway?" is a fair question, and I hear it often because I build APIClaw, a hosted flat-rate gateway. The short answer: the proxy is free, but the tokens never are. The real comparison is between four billing shapes, not between "free" and "paid". Below are the numbers from each vendor's live pricing page and a worked example you can redo with your own spend.

The four shapes

• Self-hosted LiteLLM proxy. Open source, runs on your own server. You bring provider keys and pay each provider per token. The cost is a server plus your time.
• Requesty. A hosted router. You pay the model cost plus a platform fee.
• Portkey. A hosted gateway and observability layer, priced by recorded logs. You bring your provider keys and pay those providers separately.
• APIClaw. A hosted gateway that sells model access itself, at a fixed monthly price for a daily request allowance, with no per-token billing.

Prices (prices checked Oct 2026)

Requesty's own example makes the fee concrete: a model that costs $10 per million tokens costs $10.50 through them.

A worked example

Take a small team that currently spends $400 a month on provider tokens, mostly Claude Sonnet-class models through coding agents. That's my assumption; replace it with your own invoice total.

For the first three, the bill moves with your token count. For APIClaw, it moves with your request count. That's the whole trade. If your agents send very large contexts, flat pricing is cheap per token. If your requests are small and few, tokens are cheap anyway and a monthly plan is wasted money.

Where self-hosted LiteLLM is the better choice

I'd pick LiteLLM myself in these cases:

• You need to bring your own keys. You already have negotiated rates, cloud credits, or an enterprise contract with a provider. LiteLLM routes to those keys. APIClaw doesn't accept your provider keys; it sells its own access.
• Data must stay in your network. A self-hosted proxy can run next to a VPC endpoint or a local model. No hosted gateway can match that.
• You want unlimited requests. LiteLLM has no daily cap. Its limits are whatever you configure and whatever the providers enforce.
• You have someone who enjoys running it. If a person on the team already runs Postgres and Docker in production, the maintenance cost is close to zero.

Where Portkey or Requesty are the better choice

• Portkey fits if your real problem is observability: traces, logs, prompt management, and guardrails across providers you already pay. The $49 tier is a monitoring cost, not a model cost.
• Requesty fits if you want one hosted bill across many providers, you want to pay only for what you use, and 5% is less than the time you'd spend running a proxy.

Where APIClaw fits, and where it doesn't

APIClaw makes sense when token spend is high, unpredictable, and mostly coding agents: Claude Code, Cursor, Cline-style tools that resend big contexts many times a day. A fixed $39 a month with 1,000 requests a day is easy to budget. The model's token count doesn't change the bill.

It is not the right fit if:

• You need official Anthropic or OpenAI enterprise contracts, or a BAA.
• You need unlimited requests with no daily cap. Once you're over quota, requests return an error until the 00:00 UTC reset.
• Your volume is low. A few dozen small requests a day cost less on pay-per-token.
• You rely on heavy models all day. Heavier models draw down the allowance faster than lighter ones, so check your plan's per-model figure first.

Switching is one base URL

All of these speak the OpenAI format, so moving between them is mostly a config change. If you run LiteLLM, you can even keep it and add APIClaw as one more upstream in config.yaml:

model_list:
  - model_name: flat-rate-sonnet
    litellm_params:
      model: openai/YOUR_MODEL_ID
      api_base: https://apiclaw.biz/v1
      api_key: os.environ/APICLAW_API_KEY

The openai/ prefix tells LiteLLM to treat it as an OpenAI-compatible endpoint. Without LiteLLM, point any OpenAI SDK at the base URL directly:

from openai import OpenAI

client = OpenAI(base_url="https://apiclaw.biz/v1", api_key="YOUR_API_KEY")
resp = client.chat.completions.create(
    model="YOUR_MODEL_ID",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

Take model IDs from the model hub. APIClaw has a free trial of 50 lifetime requests with no card, which is enough to check latency and compatibility before you move any traffic. Billing is crypto on the site, and other payment methods are available via support.

How I'd decide

1. Pull last month's provider invoices and your request count.
2. If you need your own keys, private networking, or no cap, self-host LiteLLM. Add Portkey if you also need deep logs.
3. If you want hosted pay-as-you-go across many providers, compare Requesty's 5% with the hours you'd spend on a proxy.
4. If your spend is dominated by coding agents and your requests fit a daily allowance, compare that number with a flat plan on apiclaw.biz.

Every option here is sensible for someone. The wrong move is picking "free" without counting the tokens and the hours.

Disclosure: I build APIClaw; this article was written with AI assistance and checked by me. Competitor prices checked Oct 2026.