FOR CLAUDE CODE · CURSOR · CLINE

Point your coding agent somewhere cheaper. Change nothing else.

One environment variable. Native Anthropic Messages API, streaming, tool calls, and prompt caching all keep working exactly as they did. Every model bills at 20% below list price, on prepaid credits — no subscription.

Claude Code

export ANTHROPIC_BASE_URL=https://globalrouterai.com
export ANTHROPIC_AUTH_TOKEN=sk-your-key

claude

Independent third-party gateway — not affiliated with Anthropic. Model behavior can differ from a first-party API. What that means

Create an API key See pricing

NO SURPRISES

20%

below list price

Every model, every token type. No tiers, no monthly minimum.

$2.48

a heavy agent hour

About 100 Sonnet 5 requests with a large cached prompt — $3.10 at list.

10%

of input for cache reads

Where most of a coding agent’s bill actually lives.

30 days

to change your mind

Unused credits refunded in full. No bonus credits, one balance.

Drop-in, not a rewrite

Native Anthropic Messages API plus an OpenAI-compatible endpoint. Change one base URL — your tool calls and streaming keep working exactly as before.

Prompt caching that actually caches

Cache reads are billed at a tenth of input. For coding agents that resend a large system prompt every turn, this is where most of the bill lives.

You can see where the money went

Per-key, per-model, per-day usage with cache hit rates. Set a daily cap on any key so a runaway loop cannot drain your balance.

MODEL	INPUT	OUTPUT	CACHED INPUT
Claude Sonnet 5RECOMMENDED	$1.60$2.00	$8.00$10.00	$0.16$0.20
Claude Opus 4.8	$4.00$5.00	$20.00$25.00	$0.40$0.50
Claude Haiku 4.5	$0.80$1.00	$4.00$5.00	$0.080$0.10

Per million tokens, in USD. Struck-through figures are list price. Full pricing

STRAIGHT ANSWERS

The things you would otherwise have to find out the hard way.

Is this an official Anthropic or OpenAI endpoint?: No. We are an independent third-party gateway and route your requests through upstream providers. Model behavior — including system-level instructions and how the model describes itself — can differ from a first-party API. The specifics
What happens when upstream capacity is degraded?: You get a 503 with a Retry-After header — never a bare 500 or a silently truncated stream. That distinction matters: coding agents retry on 503 and give up on 500. Failed requests are not billed. Live status
What is it actually tuned for?: Coding agent workloads — long sessions, a large cached system prompt resent every turn, heavy tool use. That is the traffic we test against. Evaluate it for your use case before you depend on it.
Can I get my money back?: Unused credits are refunded in full within 30 days. We give no bonus credits and run no promotions, so there is never an argument about which balance gets spent first — there is only one.